Little Pi's Tech Music Tips ｜ 小皮的聲波速記
Let AI mascot Little Pi give you a fast introduction to the world of electroacoustic music!
讓 AI 吉祥物「小皮」帶你快速認識電聲音樂的世界！
From composers, works, and sound tools to music technology, conferences, and the latest trends, Little Pi delivers one topic worth knowing in about 2–3 minutes of English narration per episode.
從作曲家、作品與聲音工具，到音樂科技、研討會與最新趨勢，小皮每集用約 2～3 分鐘的中文口白，帶你掌握一個值得知道的主題。
Whether you're a composer, musician, student, or a first-time listener to electroacoustic music, join Little Pi for a relaxed, concise journey through the endless possibilities where sound meets technology.
不論你是作曲家、音樂人、學生，還是第一次接觸電聲音樂的聽眾，都可以跟著小皮，用輕鬆、簡潔的方式，探索聲音與科技交會的無限可能。
Electroacoustic music runs deep — Little Pi makes it simple.
電聲音樂很深，小皮幫你說得簡單。
Episodes全部集數
串流時代，黑膠為什麼又紅了？黑膠唱片怎麼發聲、真的比較好聽嗎
手機就能聽上億首歌，卻有人排隊買黑膠？黑膠唱片是一條很長的螺旋溝槽，牆壁刻成聲波的形狀，唱針沿著溝槽抖動、唱頭把抖動變成電訊號。歷史：1877 年愛迪生的錫箔圓筒留聲機、1887 年柏林納改成圓盤、1948 年哥倫比亞唱片推出一面約 20 分鐘的長時間黑膠、1982 年 CD 登場後黑膠退場。2008 年美國獨立唱片店發起 Record Store Day；2022 年美國黑膠銷售張數超越 CD，是 1987 年以來第一次；2024 年美國黑膠約 4400 萬張、CD 約 3300 萬張，黑膠營收約 14 億美元、1984 年以來最高。黑膠比較好聽嗎？只比規格，CD 雜訊更少、動態更大；黑膠有爆豆聲、內圈容易失真，「溫暖感」一部分來自細微失真，加上黑膠要另做母帶（低音集中在中間，唱針才不會跳），有時也比較不被壓得那麼大聲——是聽起來不一樣，不是誰比較好。大封面、放唱片翻面的儀式感、支持歌手也是魅力；Luminate 2023 年調查，過去一年買過黑膠的人約只有一半說家裡有唱盤。
Why Is Vinyl Back in the Streaming Era? How Records Make Sound, and Do They Sound Better?
We can stream over a hundred million songs, yet people line up for vinyl. A record is one long spiral groove whose walls are cut in the shape of the sound wave; the needle wiggles along it and the cartridge turns that motion into an electrical signal. History: Edison's tin-foil cylinder phonograph (1877), Emile Berliner's flat disc (1887), Columbia's long-playing record with about 20 minutes per side (1948), and the CD (1982), after which vinyl faded. In 2008, independent US record stores launched Record Store Day; in 2022, vinyl outsold CDs by units in the US for the first time since 1987; in 2024 the US sold about 44 million vinyl records versus 33 million CDs, with vinyl revenue around 1.4 billion dollars, the highest since 1984. Does vinyl sound better? On specs, CDs have less noise and more dynamic range; vinyl has crackles and inner-groove distortion. Its warmth partly comes from subtle distortion and from separate vinyl masters (bass kept centered so the needle stays in the groove, sometimes less loudness compression). It sounds different, not better. Big covers, the ritual of playing a record, and supporting artists add to the appeal; a 2023 Luminate survey found only about half of recent US vinyl buyers said they owned a record player.
耳機開多大聲會傷耳朵？60/60 原則與 WHO 安全聆聽建議
搭捷運越開越大聲，下車耳朵嗡嗡叫？聲音大小用分貝表示：安靜圖書館約 30、一般說話約 60、很吵的車廂或馬路約 80、演唱會前排可以超過 110 分貝；每多 10 分貝能量變十倍、聽起來約大聲兩倍。世界衛生組織建議大人每週在 80 分貝以內總共聽 40 小時（小朋友 75 分貝），每多 3 分貝安全時間就少一半：86 分貝每週只剩 10 小時、92 分貝只剩 2.5 小時。內耳耳蝸裡像小草一樣的毛細胞負責把振動變成神經訊號，被吵太久會受傷，而且人類的毛細胞壞了不會再生；耳朵悶悶的、嗡嗡叫就是警訊。WHO 估計 12～35 歲年輕人約一半有聽力受損風險，2022 年 BMJ Global Health 研究估計全球約 6.7 億到 13.5 億年輕人有不安全的聆聽習慣。保護耳朵三招：60/60 原則（音量六成以下、聽六十分鐘休息）、打開手機的耳機音量提醒（iPhone 以 80 分貝、7 天 40 小時為標準）、吵的地方用降噪耳機（第 50 集）。
How Loud Is Too Loud for Headphones? The 60/60 Rule and WHO Safe Listening
Turning up your headphones on a noisy train and leaving with ringing ears? Loudness is measured in decibels: a quiet library is about 30, normal conversation about 60, a noisy train or busy road about 80, and the front rows of a concert can exceed 110. Every 10 more decibels means ten times the energy and sounds roughly twice as loud. The World Health Organization recommends that adults listen at up to 80 decibels for a total of 40 hours a week (75 decibels for children), and every 3 more decibels halves the safe time: 10 hours at 86 decibels, just 2.5 hours at 92. Hair cells in the cochlea, like tiny blades of grass, turn vibrations into nerve signals; too much loud sound damages them, and human hair cells don't grow back. Muffled or ringing ears are a warning sign. WHO estimates about half of people aged 12 to 35 are at risk, and a 2022 BMJ Global Health study estimated 0.67 to 1.35 billion young people have unsafe listening habits. Three tips: the 60/60 rule, your phone's headphone level alerts (iPhone uses 80 decibels for 40 hours over 7 days), and noise-cancelling headphones in loud places (episode 50).
閉著眼睛也知道蚊子在哪？兩隻耳朵怎麼聽出方向｜空間聲音系列①
空間聲音系列第一集！半夜蚊子在耳邊嗡嗡叫，閉著眼睛也知道在左邊還是右邊，耳朵是怎麼聽出方向的？兩隻耳朵就像頭兩側的兩支麥克風，大腦比較兩份聲音的差別：一是時間差，聲音從左邊來會先到左耳，兩耳只隔十幾公分，最多只差約 0.6～0.7 毫秒，大腦卻分辨得出約十萬分之一秒；二是音量差，頭會擋出「影子」，高音聲波短被擋住，低音聲波長繞得過去。英國物理學家雷利勳爵 1907 年提出雙重理論：低音主要靠時間差、高音主要靠音量差，分界約 1500 赫茲（他也因發現氬氣得過諾貝爾物理獎）。這也解釋了家裡低音喇叭為什麼擺角落也聽不出方向，不過房間還是會影響低音飽滿度。只靠兩耳差別有個漏洞：左前方和左後方的差別幾乎一樣，這些會搞混的位置圍成甜筒形，叫混淆錐；轉一下頭就能幫忙判斷（Wallach 1940），耳廓也是分辨前後上下的祕密武器。最後介紹定位高手倉鴞：兩個耳孔一高一低，連上下都聽得出來。
Hear a Mosquito with Your Eyes Closed? How Two Ears Find Direction | Spatial Sound Series 1
The first episode of the Spatial Sound Series! A mosquito buzzes by your ear at night, and you know which side it's on without opening your eyes. How? Your two ears work like two microphones, and the brain compares them. First, the time difference: a sound from the left reaches the left ear first; with ears only a little more than ten centimeters apart, the biggest gap is about 0.6 to 0.7 milliseconds, yet the brain can detect around a hundred-thousandth of a second. Second, the loudness difference: the head casts an acoustic shadow that blocks short high-frequency waves, while long low-frequency waves bend around it. Lord Rayleigh's duplex theory (1907) says low sounds are located mainly by time and high sounds mainly by loudness, with the dividing line around 1500 hertz (he also won the Nobel Prize in Physics for discovering argon). That's why you can hardly tell where a subwoofer sits, though the room still shapes how full the bass sounds. The loophole: front-left and back-left give almost the same differences. These confusable positions form an ice-cream-cone shape called the cone of confusion; turning your head helps (Wallach, 1940), and the outer ear is another secret weapon. Finally, the barn owl, whose ear openings sit at different heights so it can hear up and down too.
歌手沒唱過的歌，怎麼冒出來的？AI 歌聲合成與 AI 翻唱｜DiffSinger、RVC
網路上的 AI 翻唱，明明是歌手沒唱過的歌，聽起來卻超像本人？AI 唱歌有兩條路：歌聲合成是給歌詞、音符和每個音唱多久，從零生出歌聲；歌聲轉換是給一段錄好的歌，旋律歌詞不變、只換嗓子。2004 年 Yamaha 的 VOCALOID 用真人錄好的音節像拼貼一樣接起來（初音未來就是這樣唱的），接縫有機器味；現在改由神經網路直接畫出整段聲音。DiffSinger（AAAI 2022）用第 38 集的擴散模型，但想了捷徑「淺擴散」：先用簡單模型畫一張有點糊的草稿，只加一點雜訊、從中間開始擦，步數少、細節更清楚，畫好的梅爾頻譜再交給第 53 集的聲碼器；社群維護的 OpenVPI 版本可搭配免費的 OpenUtau 編輯器。RVC（檢索式語音轉換，2023 年爆紅）把「唱什麼」和「誰在唱」拆開：用第 49 集的 HuBERT 抓出唱了哪些字、RMVPE 抓音高，再從目標歌手的錄音裡檢索最像的聲音片段混進去讓音色更像，最後由解碼器用新嗓子唱出來，約 10 分鐘乾淨人聲就能練。AI 翻唱三步驟：聲源分離（第 35 集）拿出人聲 → RVC 換嗓子 → 跟伴奏混回去。最後談聲音是誰的：2023 年〈Heart on My Sleeve〉模仿 Drake 與 The Weeknd 被下架；Holly Herndon 2021 年推出 Holly+，Grimes 2023 年推出 Elf.Tech、版稅五五分。想用別人的聲音，要先經過本人同意。
Songs the Singer Never Sang? AI Singing Synthesis and AI Covers: DiffSinger and RVC
AI covers online sound just like real singers performing songs they never sang. There are two roads for AI singing: singing voice synthesis creates a voice from lyrics, notes and durations, while singing voice conversion keeps an existing recording's melody and lyrics but swaps the voice. Yamaha's VOCALOID (2004) stitched together syllables recorded by a real singer, the way Hatsune Miku sings, but the seams sound robotic; today neural networks paint the whole sound directly. DiffSinger (AAAI 2022) uses the diffusion models from episode 38 with a shortcut called shallow diffusion: a simple model draws a blurry rough draft, a little noise is added, and cleaning starts from the middle, so it takes fewer steps and gives clearer detail; the mel spectrogram then goes to a vocoder (episode 53). The community OpenVPI version works with the free OpenUtau editor. RVC (Retrieval-based Voice Conversion, viral in 2023) separates what is sung from who sings it: HuBERT (episode 49) captures the words, RMVPE captures pitch, and a retrieval step blends in the most similar snippets from the target singer's recordings before a decoder sings in the new voice; about ten minutes of clean vocals can train a model. An AI cover takes three steps: source separation (episode 35), RVC, and mixing back with the accompaniment. Finally, whose voice is it? In 2023 Heart on My Sleeve imitated Drake and The Weeknd and was taken down; Holly Herndon launched Holly+ in 2021, and Grimes launched Elf.Tech in 2023 with a fifty-fifty royalty split. Always get permission before using someone else's voice.
現場把鼓聲變人聲？筆電就跑得動的 AI 音色轉換｜IRCAM RAVE、nn~、Neutone
打一段鼓，喇叭傳出來卻像人聲？IRCAM 的 Antoine Caillon 與 Philippe Esling 在 2021 年提出 RAVE（Realtime Audio Variational autoEncoder），上一集 DDSP 一次只能處理一個音，RAVE 鼓聲、雜訊、和弦都吃。骨架是第 45 集的自編碼器：編碼器把每一小段聲音壓成幾個數字，解碼器只聽過人聲，就會把一切都用人聲說出來，留下鼓的節奏、換掉音色。兩階段訓練：先學壓縮與還原（比頻譜，但聲音有點糊），再把編碼器固定、請評審挑毛病做對抗式微調，跟第 53 集的神經聲碼器同一招。多頻帶分解把聲音切成 16 個頻帶同時算，48 kHz 高音質在一般筆電 CPU 上比即時快約 20 倍（作者實測）。上台用 nn~ 在 Max/MSP、Pure Data 裡載入模型（要匯出串流模式，否則有喀啦聲），日本 Qosmo 的 Neutone 外掛也能在編曲軟體使用，輕量版甚至能在樹莓派上即時運作。還能直接推動潛空間的數字做出新聲音、不給輸入也能自己發聲。最後比較 DDSP（一次一個音、好控制）與 RAVE（什麼聲音都吃、但沒有明確的音高旋鈕）。
Turning Drums into Voices Live? AI Timbre Transfer on a Laptop: IRCAM RAVE, nn~ and Neutone
Play some drums and out comes something like a human voice? In 2021, Antoine Caillon and Philippe Esling at IRCAM introduced RAVE (Realtime Audio Variational autoEncoder). Where DDSP handles one note at a time, RAVE takes drums, noise and chords. Its backbone is the autoencoder from episode 45: the encoder squeezes each slice of sound into a few numbers, and a decoder that has only heard voices answers everything in a voice, keeping the drum rhythm while swapping the timbre. Training has two stages: first learn to compress and restore (matching spectra, but a bit blurry), then freeze the encoder and fine-tune adversarially with a judge, the same trick as the neural vocoders in episode 53. A multi-band decomposition splits sound into 16 bands computed at once, so 48 kHz audio runs about 20 times faster than real time on an ordinary laptop CPU (the authors' own test). On stage, nn~ loads models in Max/MSP and Pure Data (export in streaming mode to avoid clicks), Qosmo's Neutone plugin runs RAVE models in a DAW, and a lightweight version even runs on a Raspberry Pi. You can also push the latent numbers directly to make new sounds, or let it generate with no input. Finally, DDSP (one note, easy control) vs RAVE (any sound, but no clear pitch knob).
唱歌變小提琴？讓 AI 轉合成器的旋鈕｜DDSP 可微分數位訊號處理、Tone Transfer、DDSP-VST
對著麥克風哼一段，出來卻是小提琴？Google Magenta 的 DDSP（可微分數位訊號處理，ICLR 2020）不讓 AI 從零畫波形，而是讓 AI 轉合成器的旋鈕。第 53 集的神經聲碼器一秒要畫好幾萬個點、模型大又吃資料；DDSP 先用 CREPE 聽出音高、再聽出大小聲，這兩條線記下「演奏什麼」，再交給只聽過小提琴的 AI 決定旋鈕怎麼轉，所以唱歌或長笛都能變成小提琴。合成器有三個零件：泛音（好幾十條正弦波，也就是加法合成，每條多大聲決定音色）、雜訊（氣音、弓擦弦）、殘響（在什麼房間演奏）。「可微分」代表每個旋鈕轉一點點、聲音差多少都算得出來，訓練時比頻譜、一路往回推，AI 就知道該往哪轉。模型比同類小好幾倍，官方範例每個不到 13 分鐘錄音就能訓練，音高和大小聲可分開改；限制是一次只能一個音。2020 年 10 月推出 Tone Transfer 網頁（長笛、薩克斯風、小號、小提琴），2022 年推出免費 DDSP-VST 外掛，錄 10–20 分鐘單音演奏就能訓練自己的音色。
Singing into a Violin? Letting AI Turn the Synth Knobs: DDSP, Tone Transfer and DDSP-VST
Hum into a microphone and out comes a violin? Google Magenta's DDSP (Differentiable Digital Signal Processing, ICLR 2020) doesn't make the AI draw waveforms from scratch; it lets the AI turn a synthesizer's knobs. Neural vocoders (episode 53) draw tens of thousands of points per second and need big models and lots of data. DDSP first extracts pitch with CREPE and loudness, the two lines describing what is played, then hands them to an AI that has only heard the violin, which decides how to turn the knobs, so a voice or a flute becomes a violin. The synthesizer has three parts: harmonics (dozens of sine waves, additive synthesis, whose levels set the timbre), filtered noise (breath, bow scratch), and reverb (the room). Differentiable means the effect of every tiny knob turn can be calculated, so training compares spectra and traces back which way to turn each knob. The models are several times smaller than similar neural synthesizers, each official example trained on under thirteen minutes of audio, and pitch and loudness can be edited separately; the limit is one note at a time. Magenta launched the Tone Transfer web app in October 2020 (flute, saxophone, trumpet, violin) and the free DDSP-VST plugin in 2022; ten to twenty minutes of single-note playing can train your own timbre model.
AI 怎麼「讀譜」？把音樂切成一個個字｜符號音樂 token、REMI、ABC 記譜法、NotaGen
之前介紹的 MusicGen 直接生出聲音，音樂 AI 還有另一派：先寫譜。語言模型讀「字」，音樂要切成什麼字？聲音 token（第 36 集）一秒約 200 個、記錄聲音本身；符號 token 記「演奏什麼」：哪個音、第幾拍、持續多久，一小節十幾個字，好改、換什麼聲音都行，但沒有音色。第一招：Google Magenta 2018 年照 MIDI 事件寫（按下、等幾秒、放開），適合記真人演奏，但時間用秒，AI 不知道拍子在哪。第二招：台灣楊奕軒老師團隊 2020 年的 REMI（Pop Music Transformer），加入「小節」和「第幾格」，一小節切 16 格像方格紙，再記音高、力度、持續幾拍，生成的流行鋼琴曲節奏穩多了。第三招：ABC 記譜法，Chris Walshaw 1993 年為記民謠整理的純文字寫譜法，語言模型天生會讀。2025 年北京中央音樂學院等團隊的 NotaGen 把 ABC 用到極致：約 160 萬首 ABC 譜預訓練、約 9 千首精選古典微調，用「年代＋音樂家＋編制」當提示寫出整份五線譜；再用 CLaMP-DPO 請 CLaMP 當評審，像真譜的多學、不像的少學，不用真人評分，盲測贏過其他模型。
How Does AI "Read" Music? Cutting Music into Words: Symbolic Tokens, REMI, ABC Notation and NotaGen
MusicGen generates sound directly, but another camp of music AI writes a score first. Language models read words, so what words should music be cut into? Audio tokens (episode 36) run about 200 per second and record the sound itself; symbolic tokens record what to play: which note, which beat, how long, about a dozen per bar, easy to edit and playable on any instrument, but with no timbre. Approach one: Google Magenta (2018) wrote MIDI events (press, wait, release), great for real performances, but time is in seconds, so AI can't find the beat. Approach two: REMI from Yi-Hsuan Yang's team in Taiwan (Pop Music Transformer, 2020) adds bar and position words, sixteen slots per bar like graph paper, plus pitch, velocity and duration, giving much steadier rhythm. Approach three: ABC notation, a plain-text notation Chris Walshaw formalized in 1993 for folk tunes, which language models read naturally. NotaGen (2025), led by the Central Conservatory of Music in Beijing, pretrains on about 1.6 million ABC scores, fine-tunes on about 9,000 classical pieces, and is prompted with period, composer and instrumentation; CLaMP-DPO then uses the CLaMP model as a judge so it learns more from outputs that resemble real scores, with no human grading, and it beat other models in blind tests.
聽一首歌就寫出音符？AI 自動採譜怎麼運作｜Basic Pitch、MT3
聽一首歌就把每個音寫下來叫採譜，以前靠訓練有素的耳朵，現在 AI 也會了。單音好辦（像第 57 集 Auto-Tune 找音高），但和弦的泛音全疊在一起，AI 容易把泛音當成另一個音或猜錯八度，還要知道每顆音何時開始、何時結束。大部分採譜 AI 先把聲音畫成頻譜圖（第 53 集），再畫出鋼琴捲簾：每條橫槓是一顆音符。Google 2018 年的 Onsets and Frames 同時練「抓起音」和「判斷還在不在響」，抓到開頭才算新音符，鋼琴採譜準確度大躍進。Spotify 2022 年開源的 Basic Pitch 輕巧不挑：人聲、吉他、鋼琴都能轉，筆電和網頁都能跑，還記下滑音和顫音，但最擅長一次一種聲音。Google Magenta 的 MT3 把採譜當翻譯：頻譜進、一串像 MIDI 的指令出（換樂器、時間、音符開關），像第 37 集的接龍一個個寫出來，多種樂器一起採還標出誰演奏。但 MIDI 還不是五線譜：要對齊拍子、決定升 Fa 還是降 Sol、分聲部、加記號，還有好幾關。
Writing Down Every Note from a Recording? How AI Music Transcription Works: Basic Pitch and MT3
Writing down every note you hear is called transcription; it used to take a trained ear, and now AI can do it. A single voice is easy (like Auto-Tune's pitch tracking in episode 57), but in a chord the overtones pile up, so AI may mistake an overtone for a note or guess the wrong octave, and it must also know when each note starts and ends. Most transcription AIs draw the sound as a spectrogram (episode 53) and then draw a piano roll, one bar per note. Google's Onsets and Frames (2018) trains two detectors at once, one for note starts and one for whether a note is still sounding, which greatly improved piano transcription. Spotify's open-source Basic Pitch (2022) is light and instrument-agnostic: voice, guitar, piano; it runs on a laptop or in the browser and captures slides and vibrato, though it works best on one source at a time. Google Magenta's MT3 treats transcription as translation: spectrogram in, a stream of MIDI-like tokens out (instrument, time, note on/off), written one by one like the word chain from episode 37, transcribing many instruments at once and labeling who played each note. But MIDI is not yet a score: aligning to beats, choosing F sharp or G flat, splitting voices and adding markings are still hurdles ahead.
為什麼流行歌的人聲都好準？Auto-Tune 修音怎麼運作
流行歌的人聲為什麼每個音都那麼準？一跳一跳的機器人聲又是怎麼來的？Auto-Tune 的發明人 Andy Hildebrand 原本是石油公司 Exxon 的工程師，用震波分析地底找石油；據說一次晚餐朋友開玩笑要他做一台讓人唱歌不走音的盒子，他發現分析震波的數學也能分析人聲，1997 年 Auto-Tune 誕生。修音分三步：先用自相關（把波形跟自己錯開比對，錯開剛好一圈時疊得最整齊）聽出你唱多高，再找出這首歌會用到的音裡最近的那一個，然後只把音高推過去、不改速度和歌詞。關鍵旋鈕「修正速度」設慢很自然，設成 0 音高會瞬間跳格子，就是 1998 年 Cher〈Believe〉的機器人聲（據說製作人一開始還說是聲碼器做的）。為什麼不會變花栗鼠？好的修音只移音高、保留共振峰（喉嚨和口腔形狀造成的嗓音特色）；Melodyne 更把每個音畫成色塊讓錄音師拖拉。T-Pain 把機器人聲變成招牌，2009 年 Jay-Z 發歌〈D.O.A.（Death of Auto-Tune）〉，但修音沒有消失：現在輕修音幾乎是錄音標配，明顯的機器人聲則成了一種效果，像電吉他的破音。
Why Do Pop Vocals Sound So Perfectly in Tune? How Auto-Tune Works
Why does every note in pop vocals sound so perfect, and where does that jumpy robot voice come from? Auto-Tune's inventor, Andy Hildebrand, was an Exxon engineer who analyzed seismic waves to find oil. As the story goes, a friend at dinner joked that he should build a box to make her sing in tune; he realized the same math could analyze the voice, and Auto-Tune arrived in 1997. It works in three steps: autocorrelation (comparing the waveform with a shifted copy of itself until it lines up after exactly one cycle) finds how high you're singing; it picks the nearest note the song uses; then it pushes only the pitch there, keeping your timing and lyrics. The key knob, retune speed, sounds natural when slow; set to zero, the pitch snaps instantly, creating the robot voice of Cher's 1998 hit Believe (the producers reportedly first claimed it was a vocoder). Why no chipmunk effect? Good pitch correction keeps the formants, the resonances shaped by your throat and mouth. Melodyne goes further, drawing each note as a draggable blob. T-Pain made the robot voice his signature, Jay-Z released D.O.A. (Death of Auto-Tune) in 2009, yet it never went away: gentle correction is now standard in recording, and the obvious robot voice is an effect, like guitar distortion.
老師的作品 11 月連跑三站！米蘭、厄瓜多、明圖爾諾｜小皮追星日記 #2
小皮追星日記第二集！林哪位老師的長笛與電聲作品《夢行者：無垠的朝聖》（Dream Voyager: A Pilgrim of the Infinite，長笛家吳正宇演奏）十一月連跑三站，老師本人無法到場，讓作品替老師出國。第一站：11/13–15 義大利米蘭 NOIS//E 2026 Festival and Symposium 的 AudioVisual Room 單元，實驗與電子音樂節慶＋研討會，免費入場。第二站：國際電聲音樂展 MUSLAB 的 KARMA 2026（KARMA＝佛教說的業力，探討因果），11/26 在厄瓜多曼塔博物館開展、活動到 12/1；MUSLAB 也是老師的老朋友：2023 年《極光》（Aurora）收錄於 Planeta Complejo 唱片製作、2024 年《冰川》（Glacier）參展、2026 年 3–5 月《極光》再度受邀於墨西哥市數位文化中心展出。第三站：11/27–28 義大利明圖爾諾 ECHI ELETTROACUSTICI 2026 國際作曲比賽，決賽以公開音樂會形式進行，28 件決賽作品現場演出，老師是唯一來自台灣的決賽作品。
The Teacher's Work Tours Three Cities in November! Little Pi's Fan Diary #2
Little Pi's Fan Diary #2! The teacher's piece for flute and electronics, Dream Voyager: A Pilgrim of the Infinite (performed by flutist Cheng-Yu Wu), makes three stops this November, traveling abroad on the teacher's behalf. Stop one: November 13-15, the AudioVisual Room programme at NOIS//E 2026 Festival and Symposium in Milan, Italy, a festival and symposium of experimental and electronic music with free entry. Stop two: KARMA 2026 by MUSLAB, the international electroacoustic music exhibition (karma, the Buddhist idea of cause and effect), opening November 26 at the museum in Manta, Ecuador, with events through December 1. MUSLAB is an old friend: Aurora was included in its Planeta Complejo record production in 2023, Glacier was shown in 2024, and Aurora was invited back to the Digital Culture Center in Mexico City from March to May 2026. Stop three: November 27-28, the ECHI ELETTROACUSTICI 2026 International Composition Competition in Minturno, Italy, whose final round is a public concert of all 28 finalist works; the teacher's piece is the only finalist from Taiwan.
這首歌是 AI 做的嗎？音樂浮水印與 AI 偵測怎麼運作
AI 做的歌越來越像真的，Deezer 2026 年 7 月公布：每天新上傳的歌已經有一半以上是 AI 做的（6 月高峰約一天 9 萬首）。分辨 AI 聲音有兩條路：浮水印是生成時就在聲音裡加入人耳聽不出的暗號，之後用解碼器讀出來，很準但要 AI 公司配合；偵測是事後找 AI 留下的痕跡，誰的歌都能查但可能猜錯。聲音浮水印用兩個神經網路一起練（嵌入器＋偵測器），中間故意轉 MP3、加雜音、加速、剪片段來搗亂。Meta 2024 開源的 AudioSeal 每 1/16000 秒打一個分數，混剪也能標出 AI 那一段，偵測比 WavMark 快最多約兩個數量級。Google 的 SynthID 藏在 Lyria 做的每段聲音裡，10/7 檢查網站開放給所有人，但只認得 Google 與合作夥伴的暗號，Suno、Udio 的歌查不到——沒偵測到不代表是真人做的。沒有浮水印怎麼辦？Deezer 團隊發現自編碼器把資料放大還原時，會在頻譜上留下規律的小尖刺，像 AI 的指紋，好幾種情況準確度超過 99%，獲 ISMIR 2025 最佳論文。但升降音高、再壓縮、換模型架構都會讓偵測變難；AI 歌雖佔一半上傳，只佔總播放 1–3%，其中約 85% 是灌播放數，Deezer 不付錢。
Was This Song Made by AI? How Audio Watermarks and AI Detection Work
AI songs sound more and more real: in July 2026 Deezer announced that more than half of the new songs uploaded each day are made by AI (about 90,000 a day at the June peak). There are two ways to tell AI audio apart. Watermarking slips an inaudible code into the sound at generation time and reads it back later with a decoder: very accurate, but it needs the AI company's cooperation. Detection looks for traces the AI left behind: it works on anyone's song but can guess wrong. Audio watermarks are trained with two neural networks together (an embedder and a detector), with deliberate attacks in between: MP3 conversion, noise, speed-ups, cuts. Meta's open-source AudioSeal (2024) scores every 1/16,000 of a second, so it can highlight the AI part of a spliced clip, and detects up to about 100x faster than WavMark. Google's SynthID sits in every sound made by Lyria; its checking site opened to everyone on October 7, but it only recognizes Google's and partners' codes, so Suno and Udio songs come back not detected. Not detected does not mean human-made. With no watermark, Deezer's team found that autoencoders leave evenly spaced little spikes on the spectrum when they scale data back up into sound, an AI fingerprint that reaches over 99% accuracy in several scenarios (ISMIR 2025 best paper). But pitch shifting, recompression, and new architectures make detection harder. AI songs are half the uploads yet only 1-3% of plays, and about 85% of those plays are fraudulent and go unpaid.
AI 不繞路了！流匹配 Flow Matching 讓生成走直線
擴散模型要一步一步擦掉雜訊，從雜訊走回聲音的路常常彎彎曲曲，所以步數多、速度慢。流匹配（Flow Matching）的想法是乾脆走直線：每次抽一段真聲音和一把雜訊連成直線，在線上隨便選一點（兩者按比例混合），考模型「該往哪走、走多快」，正確答案就是真聲音減掉雜訊，不用模擬整條路。練久了模型心中就有一張箭頭地圖（向量場），生成時從雜訊出發、跟著箭頭走幾步就到。但幾百萬條直線交叉，整體路線還是有點彎：2022 年德州大學團隊的整流（Rectified Flow）用練好的模型再配一次對、再練一次把路拉直；同年 Meta 發表 Flow Matching 論文。語音 AI 最早大量採用：Meta 的 Voicebox（2023）、上海交通大學等團隊開源的 F5-TTS（2024，10 萬小時多語言，十秒語音約一秒半做好，Sway Sampling 把步子多花在開頭）。音樂方面：開源 ACE-Step（2025）在潛空間用流匹配，A100 約 20 秒做出 4 分鐘的歌；2026 年 5 月的 Stable Audio 3 以流匹配打底，再加對抗式後訓練減少步數，H200 不到 2 秒，小型和中型開放權重。
AI Stops Taking Detours! Flow Matching Walks a Straight Line
Diffusion models wipe away noise one step at a time, and the path from noise back to sound is often twisty, so they need many steps and run slowly. Flow matching's idea is to just walk a straight line: take a piece of real sound and a handful of noise, connect them, pick any point on the line (the two mixed in proportion), and test the model on which way to go and how fast. The right answer is the real sound minus the noise, with no need to simulate the whole path. After lots of practice the model has an arrow map in its head (a vector field); to generate, start from noise and follow the arrows for a few steps. But millions of crossing lines still make the overall path bend a little: rectified flow (University of Texas, 2022) re-pairs with the trained model and trains again to straighten it, and Meta published the Flow Matching paper the same year. Speech AI adopted it first: Meta's Voicebox (2023) and the open-source F5-TTS (2024, led by Shanghai Jiao Tong University, 100,000 hours of multilingual speech, ten seconds of speech in about 1.5 seconds, with Sway Sampling spending more steps at the start). In music, open-source ACE-Step (2025) uses flow matching in latent space and makes a four-minute song in about 20 seconds on an A100; Stable Audio 3 (May 2026) builds on flow matching plus adversarial post-training to cut steps, taking under 2 seconds on an H200, with small and medium weights released.
AI 先畫照片再出聲？梅爾頻譜與神經聲碼器
很多會說話、會唱歌的 AI 不是直接做出聲音，而是先畫一張「聲音的照片」——梅爾頻譜，再請神經聲碼器把照片變回聲音。頻譜圖橫軸是時間、縱軸是高低、越亮越大聲；梅爾刻度（1937 年提出，名字來自 melody）照耳朵的感覺畫尺，低音區切得細、高音區切得粗，常用 80 格。但頻譜圖丟掉了相位（每條波的起跑點），1984 年的 Griffin-Lim 靠反覆猜測補回來，2022 年爆紅的 Riffusion 就用這招，聲音常帶金屬味。神經聲碼器讓 AI 學著畫波形：2016 年 DeepMind 的 WaveNet 一個點一個點畫，好聽但很慢；2020 年韓國 Kakao 的 HiFi-GAN 用生成對抗網路，仿畫家照梅爾頻譜畫波形、多組鑑定師分辨真假，在一張 V100 GPU 上比即時快 167.9 倍；2022 年 NVIDIA 的 BigVGAN 是通用聲碼器，沒聽過的歌聲、小提琴也畫得出來，第二版支援 44.1kHz。AudioLDM 先用擴散模型畫梅爾頻譜再交給 HiFi-GAN 出聲；MusicGen 則走 EnCodec 解碼器的另一條路。
AI Draws a Photo Before It Speaks? Mel Spectrograms and Neural Vocoders
Many AIs that talk or sing don't make sound directly: they first draw a "photo of sound," the mel spectrogram, and a neural vocoder turns it back into audio. A spectrogram puts time on the horizontal axis and pitch on the vertical, with brighter meaning louder; the mel scale (proposed in 1937, named after "melody") redraws the ruler the way ears hear, with fine slices for low sounds and coarse ones for high, often just 80 bands. But a spectrogram throws away phase, each wave's starting line. Griffin-Lim (1984) guesses it back over and over, and viral 2022 app Riffusion used it, giving a metallic flavor. Neural vocoders let AI learn to paint waveforms: DeepMind's WaveNet (2016) painted one point at a time, beautiful but slow; Kakao's HiFi-GAN (2020) uses a generative adversarial network, a forger painting waveforms from the mel spectrogram while teams of detectives judge real from fake, running 167.9x faster than real time on one V100 GPU; NVIDIA's BigVGAN (2022) is a universal vocoder that handles unseen singing and violins, and v2 supports 44.1 kHz. AudioLDM draws a mel spectrogram with diffusion and hands it to HiFi-GAN, while MusicGen takes the other road with the EnCodec decoder.
AI 怎麼聽懂你的提示詞？CLAP 文字和聲音的媒人
在 AI 音樂生成網站打一句「下雨天的慵懶爵士鋼琴」就有音樂，AI 怎麼知道文字跟哪種聲音配得起來？主角是 CLAP（對比式語言聲音預訓練），靈感來自 OpenAI 2021 年圖片配文字的 CLIP。它像兩座塔：聲音塔和文字塔把狗叫聲與「狗在叫」各自變成 512 個數字，放進同一張地圖，意思相配就靠近。訓練用配對遊戲：一批聲音和描述排成表格，對角線的正確配對拉近、其他推遠，不用人工貼標籤。練好後能零樣本分類：LAION-CLAP（2022）用約 63 萬組資料，在 ESC-50 五十種環境聲猜中 91.0%，比微軟 CLAP 的 82.6% 高。生成 AI 也用它：AudioLDM 練習時只看聲音座標、生成時換成文字座標；Google MusicLM 用類似的 MuLan（4400 萬首錄音）；第一代 Stable Audio 用自訓的 CLAP，Stable Audio Open 改用 T5。CLAP 分數也常被拿來評 AI 成品跟提示詞配不配。弱點：2023 年研究發現它分不太出「先後」和「同時」，比較像只認得關鍵字。
How Does AI Understand Your Prompt? CLAP, the Matchmaker Between Text and Sound
Type "lazy jazz piano on a rainy day" into an AI music site and you get music, but how does AI know which sound goes with which sentence? Meet CLAP (Contrastive Language-Audio Pretraining), inspired by OpenAI's 2021 image-text model CLIP. It works like two towers: a sound tower and a text tower turn a dog's bark and the words "a dog is barking" into 512 numbers each, placed on one shared map where matching meanings sit close together. Training is a matching game: a batch of sounds and descriptions forms a table, true pairs on the diagonal are pulled together and the rest pushed apart, with no hand labeling needed. The result is zero-shot classification: LAION-CLAP (2022), trained on about 630,000 pairs, scores 91.0% on the 50-class ESC-50 environmental sound test, above Microsoft CLAP's 82.6%. Generative AI uses it too: AudioLDM trains on sound coordinates and swaps in text coordinates to generate; Google's MusicLM uses the similar MuLan (44 million recordings); the first Stable Audio used its own CLAP, while Stable Audio Open switched to T5. The CLAP score is also a common way to grade whether AI output matches its prompt. Its weakness: a 2023 study found it can't really tell "one after another" from "at the same time"; it mostly notices keywords.
聲音從頭頂飛過？杜比全景聲 Dolby Atmos 與空間音訊是什麼
電影院裡直升機好像真的從頭頂飛過，現在連串流平台和耳機也標榜杜比全景聲、空間音訊，它跟以前的環繞音效差在哪？從左右兩聲道的立體聲，到五點一、七點一環繞聲（點一是重低音喇叭），傳統做法是混音時就把每個聲音分配到固定的喇叭。杜比全景聲（Dolby Atmos）第一個升級是高度：天花板也裝上喇叭。更關鍵的發明是聲音物件：背景環境音照傳統方式固定在「聲道床」，像一隻飛過的鳥則是帶著座標的物件，播放時系統再依照現場喇叭的數量和位置，算出該由哪些喇叭出聲；一部片最多128軌，電影院最多可接64個喇叭，同一份混音大小影廳都能自己適應。2012年皮克斯動畫《勇敢傳說》（Brave）是第一部用杜比全景聲上映的電影。家裡沒有天花板喇叭？聲霸會把聲音往上打、從天花板反射下來；耳機則模擬頭和耳朵改變聲音的方式騙過大腦，加上頭部追蹤，轉頭時聲音會留在原位。2021年Apple Music推出杜比全景聲空間音訊，但混音師要另外做一版沉浸式混音，不是按個鈕舊歌就變立體。
Sound Flying Over Your Head? What Dolby Atmos and Spatial Audio Really Are
At the movies a helicopter seems to fly right over your head, and now streaming services and headphones advertise Dolby Atmos and spatial audio. How is that different from old surround sound? From two-channel stereo to 5.1 and 7.1 surround (the point one is the subwoofer), the traditional approach assigns every sound to a fixed speaker during mixing. Dolby Atmos first adds height, with speakers on the ceiling. The bigger breakthrough is the sound object: background ambience stays in a fixed channel bed, while something like a passing bird is an object carrying coordinates, and at playback the system works out which speakers should play it based on how many speakers the room has and where they are. A film can carry up to 128 tracks and a theater can drive up to 64 speakers, so one mix adapts to large and small rooms alike. Pixar's Brave (2012) was the first film released in Dolby Atmos. No ceiling speakers at home? Soundbars bounce sound off the ceiling, and headphones simulate how your head and ears change sound to fool your brain, with head tracking keeping sounds in place when you turn. Apple Music launched spatial audio with Dolby Atmos in 2021, but mixers have to create a separate immersive mix; it isn't a button that turns old songs 3D.
戴上耳機世界就安靜了？降噪耳機ANC怎麼把噪音變不見
搭捷運、坐飛機按下降噪鍵，噪音瞬間變小，原來耳機是用聲音去消滅聲音。小皮先分清楚兩種安靜：耳罩、耳塞把聲音擋在外面的被動隔音，和用麥克風加晶片做出抵銷聲音的主動降噪（ANC）。核心原理：聲音是空氣壓力一高一低的起伏，麥克風聽到噪音，晶片立刻算出上下顛倒的波，從喇叭播出，高點碰低點、低點碰高點，加起來互相抵銷，而且要算得非常快。麥克風有三種放法：放外側先聽到噪音的前饋式、放耳朵旁邊檢查殘留噪音的回饋式、裡外都放的混合式。為什麼開了降噪還聽得到人講話？引擎、冷氣這種穩定的低沉嗡嗡聲最好抵銷，人聲、鍵盤聲變化快、波浪又密，主要還是靠被動隔音來擋。歷史：1930年代德國物理學家Paul Lueg提出反相消音專利；1978年Bose創辦人Amar Bose在從歐洲飛回美國的班機上被噪音吵到，開始畫草圖；1989年推出飛行員用降噪耳機，2000年推出給乘客的QuietComfort。最後是通透模式：反過來用外側麥克風把外面的聲音放進來，走路、過馬路記得切換；而且環境安靜了，就不用把音樂開很大聲。
The World Goes Quiet When You Put Them On? How Noise Cancelling Headphones (ANC) Make Noise Disappear
On the subway or a plane, you press the noise cancelling button and the rumble suddenly drops: your headphones are using sound to destroy sound. Little Pi first separates two kinds of quiet: passive isolation, where ear cups and ear tips block sound out, and active noise cancellation (ANC), where microphones and a chip create a sound that cancels the noise. The core idea: sound is air pressure going up and down; a microphone hears the noise, the chip instantly calculates an upside-down copy of the wave and plays it, so the peaks meet the troughs and they cancel out, which has to happen extremely fast. There are three microphone designs: feedforward on the outside, feedback near your ear checking what noise is left, and hybrid using both. Why can you still hear people talking? Steady low hums like engines and air conditioners are easiest to cancel, while fast-changing sounds like speech and typing are mostly blocked by passive isolation. History: in the 1930s German physicist Paul Lueg patented cancelling noise with an inverted wave; in 1978 Amar Bose, founder of Bose, was bothered by noise on a flight from Europe to the US and began sketching; a pilots' headset launched in 1989, and QuietComfort for passengers in 2000. Finally, transparency mode uses the outside microphones the other way to let outside sound in, so switch it on when walking or crossing the street; and in a quieter environment you don't need to turn your music way up.
讓 AI 的聲音積木帶著意思！X-Codec 語意特徵｜為什麼加了語意，AI 比較不會唸錯字
第36集只用一句話帶過的X-Codec，這集小皮專門拆解。EnCodec、DAC這類編碼器為壓縮而生，只在乎還原的聲音像不像，不在乎聲音在說什麼，像只抄筆畫、不懂字義的速記員，所以VALL-E這類用音訊token合成語音的模型常跳字、唸錯字。X-Codec請來Meta在2021年提出的HuBERT：靠「遮住一段、自己猜」的自監督學習，學會分辨這是什麼聲音，也就是語意特徵。X-Codec讓聲音走兩條路：聲學編碼器記音色細節、HuBERT記這是什麼聲音，兩者在進入密碼本（RVQ）前合併；解碼時要交兩份作業：還原波形、還原語意特徵。成績：同一套VALL-E只換編碼器，字錯誤率EnCodec 6.37%、DAC 6.81%、X-Codec 3.26%，接近真人錄音2.23%；音質評分3.02升到4.22。音樂與音效版本用約20萬小時聲音自訓通用HuBERT，音樂接龍、文字生成音效都變好，代價是波形細節略差。開源歌曲模型YuE用的是X-Codec精簡版；2025年的X-Codec 2只用一本65536個號碼的密碼本、每秒50塊積木、語意改用Wav2Vec2-BERT，並做出語音合成模型Llasa（1B、3B、8B）。
Giving AI's Sound Blocks Meaning! X-Codec Semantic Features and Why Fewer Words Go Wrong
X-Codec only got one sentence in episode 36, so this episode Little Pi takes it apart. Codecs like EnCodec and DAC were built for compression: they care whether the rebuilt sound sounds the same, not what it means, like a stenographer copying strokes without understanding the words, so models like VALL-E that build speech from audio tokens often skip or misread words. X-Codec brings in HuBERT (Meta, 2021), which learns by hiding part of a clip and guessing it, and so learns to tell what a sound is: semantic features. X-Codec sends sound down two paths, an acoustic encoder for timbre and detail and HuBERT for meaning, merged before the codebook (RVQ); decoding must hand in two assignments, rebuilding the waveform and rebuilding the semantic features. Results with the same VALL-E setup: word error rate 6.37% with EnCodec, 6.81% with DAC, 3.26% with X-Codec, close to 2.23% for real recordings; speech quality score up from 3.02 to 4.22. For music and sound, a general-purpose HuBERT trained on about 200,000 hours improved music continuation and text-to-sound, at the cost of slightly less waveform detail. The open-source song model YuE uses a compact X-Codec; X-Codec 2 (2025) uses a single 65,536-entry codebook at 50 tokens per second with Wav2Vec2-BERT, powering the Llasa speech model (1B, 3B, 8B).
不靠老師，靠評審挑毛病！對抗式後訓練 ARC｜Stable Audio Open Small 怎麼變快
第47集小皮更正：Stable Audio Open Small變快主要不是靠蒸餾，而是對抗式後訓練ARC。不靠老師模型，靠一個專門挑毛病的評審。對抗的點子來自2014年Ian Goodfellow等人的生成對抗網路GAN：生成器像做假鈔、判別器像驗鈔員。研究團隊先訓練好整流模型（Rectified Flow），主體從約10.6億參數瘦身到3.4億，原本要50步，後訓練請評審特訓到只走8步。第一招相對式對抗：同一句提示的真錄音和生成作品擺在一起，評審要說哪個比較真。第二招對比式：把提示洗牌，評審要抓出配錯的提示，學會檢查貼不貼題，因此不用第42集的CFG（每步算兩次、一次有提示一次沒有、再放大差距），每步只算一次，記憶體省將近一半。成果：H100生成12秒立體聲約75毫秒、RTX 3090 187毫秒、Arm手機約6.6秒。聽眾測驗的多樣性ARC 4.4分、蒸餾派Presto! 2.7分，代價是音質和貼題度稍低。而且兩篇的第一作者都是Zachary Novack。
No Teacher, Just a Picky Judge! Adversarial Post-Training (ARC) and Stable Audio Open Small
Episode 47 corrected that Stable Audio Open Small gets its speed mainly not from distillation but from adversarial post-training (ARC): no teacher model, just a judge whose job is finding faults. The adversarial idea comes from GANs (Ian Goodfellow et al., 2014): a generator like a counterfeiter and a discriminator like a bank inspector. The team first trained a rectified flow model, slimming Stable Audio Open's core from about 1.06 billion to 340 million parameters; it normally needs 50 steps, and post-training coaches it down to 8. Trick one, relativistic adversarial training: a real recording and a generated clip for the same prompt are compared side by side, and the judge says which is more real. Trick two, contrastive training: prompts are shuffled and the judge must catch mismatches, learning to check prompt fit, so the model no longer needs CFG from episode 42 (computing every step twice, with and without the prompt, then amplifying the difference) and computes each step once, saving almost half the memory. Results: 12 seconds of stereo in about 75 ms on an H100, 187 ms on an RTX 3090, about 6.6 s on an Arm phone. In listening tests ARC scored 4.4 for variety versus 2.7 for the distillation method Presto!, at the cost of slightly lower quality and prompt matching. And both papers share the same first author, Zachary Novack.
同樣叫蒸餾，差很多！語言模型 vs 音樂 AI 的「蒸餾」｜Distillation
LLM圈很流行講「蒸餾（Distillation）」，音樂生成AI也在用，但同一個詞其實是兩件事。蒸餾來自2015年Geoffrey Hinton等人的論文：大老師模型帶小學生模型，不只給標準答案，還給「七成像小提琴、兩成像中提琴」的軟標籤，並用第41集的溫度讓提示更明顯。語言模型的蒸餾是「大變小」：2019年Hugging Face的DistilBERT小40%、快60%、保留約97%語言理解能力；2025年DeepSeek-R1用約80萬筆示範訓練出1.5B到70B的Qwen、Llama蒸餾版。音樂擴散模型的蒸餾是「多步變少步」：2022年Google的Tim Salimans與Jonathan Ho提出漸進式蒸餾，老師走兩步、學生一步走完，每輪砍一半，從8192步砍到4步；之後有一致性蒸餾；2024年Adobe與UC San Diego的Presto!連步數帶層數一起蒸餾，快10到18倍，32秒單聲道音樂只要0.23秒。最後小皮更正：第46集說Stable Audio Open Small靠蒸餾，其實它主要用的是對抗式後訓練ARC，作者強調不是蒸餾。
Same Word, Different Jobs! Distillation in Language Models vs Music AI
Distillation is a hot word in the language model world, and music generation AI uses it too, but the same word means two different jobs. It comes from a 2015 paper by Geoffrey Hinton and colleagues: a big teacher model guides a small student, giving not just the right answer but soft labels like seventy percent violin, twenty percent viola, using the temperature from episode 41 to make the hints clearer. Language model distillation turns big into small: Hugging Face's DistilBERT (2019) is 40% smaller, 60% faster and keeps about 97% of BERT's language understanding; DeepSeek-R1 (2025) used about 800,000 worked examples to train distilled Qwen and Llama models from 1.5B to 70B. Music diffusion distillation turns many steps into few: progressive distillation (Tim Salimans and Jonathan Ho, Google, 2022) has the student do in one step what the teacher does in two, halving each round from 8192 steps down to 4; then came consistency distillation; and Presto! (Adobe and UC San Diego, 2024) distills both steps and layers, 10 to 18 times faster, 32 seconds of mono music in 0.23 seconds. Finally, a correction: episode 46 said Stable Audio Open Small relies on distillation, but it mainly uses adversarial post-training (ARC), which its authors stress is not distillation.
擦 1000 步太慢了！擴散模型怎麼越生越快｜雜訊排程與 DDIM
第38集說擴散模型從雜訊一步步擦出音樂，但到底要擦幾步？每一步擦多少？這集小皮介紹雜訊排程（Noise Schedule）與步數（Sampling Steps）。雜訊排程是訓練時加雜訊的時間表，像把果汁慢慢加水稀釋。2020年DDPM用1000步的線性排程，但加得太快，後面一大段聲音早就不見了，等於白做工；2021年OpenAI的Alex Nichol與Prafulla Dhariwal提出餘弦排程（Cosine Schedule），頭尾放慢，每一步都有事做。生成時若也走1000步太慢，2020年Stanford團隊的DDIM讓同一個模型跨大步，論文實測快10到50倍。步數太少雜訊擦不乾淨、太多則變慢且聽不出分別，Stable Audio Open這類工具常見設定在幾十到一兩百步，還能選不同的取樣器（Sampler）。最後介紹蒸餾（Distillation）、2023年的一致性模型（Consistency Models），以及2025年Stability AI與Arm推出、官方宣稱手機8秒內能生成11秒聲音的Stable Audio Open Small。
1000 Steps Is Too Slow! How Diffusion Models Got Faster: Noise Schedules and DDIM
Episode 38 said diffusion models wipe noise away step by step until music appears. But how many steps, and how much noise per step? Little Pi explains the noise schedule and sampling steps. The noise schedule is the timetable for adding noise during training, like slowly watering down a glass of juice. In 2020 DDPM used a 1000-step linear schedule, but it adds noise too fast, so a long stretch near the end is pure noise and wasted work. In 2021 OpenAI's Alex Nichol and Prafulla Dhariwal proposed the cosine schedule, slowing down at both ends so every step counts. Generating with 1000 steps is too slow, so Stanford's DDIM (2020) lets the same model take big strides, 10 to 50 times faster in the paper's experiments. Too few steps leave grainy hiss, too many are slow with no audible gain; tools like Stable Audio Open usually use a few dozen to one or two hundred steps, with a choice of samplers. Finally: distillation, Consistency Models (2023), and Stable Audio Open Small (2025, Stability AI and Arm), officially generating 11 seconds of audio on a phone in under 8 seconds.
AI 把聲音收進哪裡？潛空間與自編碼器 Latent Space & Autoencoder
第38集說Stable Audio Open在「壓縮過的空間」裡做擴散，那個空間在哪裡？這集小皮介紹潛空間（Latent Space）與自編碼器（Autoencoder）：編碼器把聲音壓小、解碼器再還原，中間最窄的瓶頸只放得下精華，像出國用真空壓縮袋打包行李。瓶頸裡的數字就像地圖座標，相像的聲音會靠在一起——長笛和雙簧管是鄰居，鼓聲住在另一區，這是模型自己整理出來的。2017年Google Magenta的NSynth用約30萬個單音訓練，在長笛和風琴之間取中間點，就得到世界上本來不存在的新音色，還做成實體合成器NSynth Super；2018年MusicVAE能在兩段旋律之間慢慢變過去。2013年的變分自編碼器（VAE）讓每個聲音變成一團雲，地圖沒有破洞。Stable Audio Open用VAE把44.1 kHz立體聲（每秒88,200個數字）壓成每秒21.5格×64個數字，小了約64倍，在小地圖上擴散再解碼，這就是潛在擴散（Latent Diffusion）。最後比較EnCodec（瓶頸後接RVQ變號碼、適合接龍）和Stable Audio（連續座標、適合擴散）。
Where Does AI Keep Sound? Latent Space and Autoencoders Explained
Episode 38 said Stable Audio Open does diffusion inside a compressed space. Where is that space? Little Pi introduces the latent space and the autoencoder: an encoder squeezes sound down, a decoder rebuilds it, and the narrow bottleneck in the middle only fits the essentials, like packing clothes in vacuum storage bags. The numbers in the bottleneck are like map coordinates, and similar sounds end up close together: flute and oboe are neighbors, drums live in another neighborhood, and the model organizes this by itself. In 2017 Google Magenta's NSynth was trained on about 300,000 single notes; taking the point halfway between a flute and an organ yields a brand-new tone color, later built into the NSynth Super hardware synth. In 2018 MusicVAE morphed gradually between melodies. The 2013 Variational Autoencoder (VAE) turns each sound into a small cloud, so the map has no holes. Stable Audio Open uses a VAE to squeeze 44.1 kHz stereo (88,200 numbers per second) into 21.5 frames per second of 64 numbers each, about 64 times smaller, then runs diffusion on that small map and decodes it back: latent diffusion. Finally, EnCodec (RVQ codes after the bottleneck, for token prediction) versus Stable Audio (continuous coordinates, for diffusion).
AI 怎麼知道該聽哪裡？注意力機制 Attention
AI寫第9小節的時候，怎麼知道要回頭看第1小節？這集小皮拆解第37集一句帶過的注意力機制（Attention）。舊方法像傳話遊戲，資訊一格一格往後傳，越傳越忘；注意力讓每一格直接回頭看前面所有格子，打分數（加起來100%），越高分越注意。分數怎麼打：每一格準備查詢（Query）、鑰匙（Key）、內容（Value），拿查詢去對鑰匙，越對得上就讀越多內容，像在圖書館先看書背標籤再決定翻哪本。多頭注意力（Multi-Head Attention）同時用好幾組，各看各的重點，分工由模型自己學。音樂看自己叫自注意力，看文字叫交叉注意力——MusicGen先用T5把文字變成數字，再用交叉注意力回頭看。注意力2014年先用在機器翻譯，2017年《Attention Is All You Need》誕生Transformer，2018年Google Magenta的Music Transformer寫出一分鐘、前後有呼應的鋼琴音樂；代價是格子數變2倍，要比對的組合變4倍。
How Does AI Know Where to Listen? The Attention Mechanism Explained
When AI writes bar nine, how does it know to look back at bar one? Little Pi unpacks the attention mechanism, mentioned briefly in episode 37. Older models worked like a game of telephone, passing information along step by step and forgetting more the farther it went. Attention lets every step look straight back at all earlier steps and give each a score (adding up to 100%): the higher the score, the more attention. How are scores decided? Each step prepares a Query, a Key and a Value; the query is matched against every key, and the better the match, the more of that value gets read, like checking book spines in a library before deciding what to open. Multi-head attention runs several sets at once, each learning its own focus. Music looking at itself is self-attention; looking at text is cross-attention: MusicGen turns the prompt into numbers with T5, then glances back at it with cross-attention. Attention was first used for machine translation in 2014; the 2017 paper Attention Is All You Need gave birth to the Transformer; and in 2018 Google Magenta's Music Transformer wrote minute-long piano music with themes that answer each other. The price: double the steps, four times the comparisons.
AI 的聲音密碼本怎麼越疊越清楚？RVQ 殘差向量量化 Residual Vector Quantization
AI把聲音變成密碼時，為什麼要用好幾本密碼本？這集小皮拆解第36集一句帶過的RVQ（Residual Vector Quantization，殘差向量量化）：編碼器把聲音切成小格、每格變成一串數字（向量），去密碼本找最接近的號碼。只用一本不是太粗糙、就是大到存不下——SoundStream論文的例子：3 kbps、每秒100格，一本要約10億個號碼；改成5本、每本64個，只要320個。RVQ讓每一本只補上一本剩下的差距（殘差），像付錢先用百元鈔、再用十元、一元。EnCodec每秒75格、每本1024個號碼，2本＝1.5 kbps、32本＝24 kbps；MusicGen用4本、每秒50格。SoundStream訓練時隨機只用前幾本（量化器丟棄Quantizer Dropout），一個模型就能選多種音質，也讓第一本最關鍵。多級向量量化最早可追溯到1982年Juang與Gray的語音編碼研究。
How Do AI Audio Codebooks Get Clearer Layer by Layer? Residual Vector Quantization (RVQ) Explained
Why does AI use several codebooks to turn sound into tokens? Little Pi unpacks RVQ, Residual Vector Quantization, mentioned briefly in episode 36. The encoder cuts sound into short slices, turns each into a vector, and looks up the closest entry in a codebook. One codebook is either too rough or too huge: in the SoundStream paper's example, at 3 kbps and 100 slices per second, a single codebook would need about a billion entries, while 5 codebooks of 64 entries need only 320. In RVQ each codebook only fixes what the previous one left over (the residual), like paying with hundreds, then tens, then ones. EnCodec uses 75 slices per second and 1024 entries per codebook: 2 codebooks = 1.5 kbps, 32 = 24 kbps; MusicGen uses 4 codebooks at 50 slices per second. SoundStream's quantizer dropout randomly uses only the first few codebooks during training, so one model supports multiple quality levels and the first codebook carries the most important information. Multi-stage vector quantization goes back to Juang and Gray's 1982 speech coding work.
AI 為什麼聽得懂提示詞？CFG 無分類器引導 Classifier-Free Guidance
明明打了「爵士薩克斯風」，AI卻生出一段鋼琴？這集小皮介紹讓AI乖乖聽話的旋鈕CFG（Classifier-Free Guidance，無分類器引導）。舊做法「分類器引導」（Dhariwal與Nichol，2021）要另外訓練一個評審模型；2021年Jonathan Ho與Tim Salimans改成同一個模型每一步算兩次：一次有提示詞、一次沒有，再把兩者的差距放大（沒提示＋強度×（有提示−沒提示））。訓練時隨機拿掉提示詞，MusicGen就是20%的時間不給提示詞，一個模型身兼兩職。強度怎麼設定：MusicGen預設3（cfg_coef），Stable Audio Open官方範例用7；太高只剩一種聲音、容易失真，而且每步算兩次，生成時間約變兩倍。
Why Does AI Follow Your Prompt? Classifier-Free Guidance (CFG) Explained
You type "jazz saxophone" and get a piano piece? Little Pi introduces CFG, Classifier-Free Guidance, the knob that makes AI follow your prompt. The older classifier guidance (Dhariwal and Nichol, 2021) needed a separate judge model. In 2021, Jonathan Ho and Tim Salimans had the same model calculate twice at every step, once with the prompt and once without, then amplify the difference (unconditional + scale x (conditional - unconditional)). During training the prompt is randomly dropped: MusicGen drops it 20% of the time, so one model does both jobs. How high to set it: MusicGen defaults to 3 (cfg_coef), and Stable Audio Open's official example uses 7. Too high leaves only one sound and can distort, and calculating twice roughly doubles generation time.
AI作曲也在擲骰子？取樣溫度 Temperature 與 top-k、top-p
同一句提示詞按兩次，為什麼生出兩首不一樣的歌？這集小皮拆解AI作曲背後的骰子：自回歸模型每一步給的是一張候選清單，每個候選各有一個百分比（MusicGen每本密碼本有1024種候選）。每次都挑第一名（貪婪解碼）會平淡又一直打轉，所以改成照百分比抽籤。溫度Temperature把分數除以溫度再換成百分比：低溫保守、高溫冒險；top-k只留前k名再抽（MusicGen預設k=250）；top-p（核取樣，Holtzman等人2019）從第一名累加到p就停，名額會自己伸縮。MusicGen預設溫度1、top-k 250、top-p 0（代表不用）。
Is AI Rolling Dice When It Composes? Temperature, Top-k and Top-p
Same prompt, two different songs? Little Pi takes apart the dice behind AI music generation. At every step, an autoregressive model gives a list where every candidate has a probability (each MusicGen codebook has 1,024 candidates). Always picking number one (greedy decoding) is bland and repetitive, so models sample instead. Temperature divides the scores before turning them into probabilities: low is safe, high is risky. Top-k keeps only the top k candidates (MusicGen's default is 250). Top-p, or nucleus sampling (Holtzman et al., 2019), adds probabilities from the top until it passes p, so the number of candidates stretches and shrinks. MusicGen's defaults: temperature 1, top-k 250, top-p 0 (off).
讓旋律自己演化？基因演算法 Genetic Algorithm 作曲
旋律也會演化？這集小皮介紹基因演算法（Genetic Algorithm）：1975年John Holland把達爾文的「適者生存」寫成演算法。先隨機生出一群旋律，每個音就是一個基因，幫每段旋律打分數（適應度）；挑高分的當爸媽，剪一刀前半接後半（交叉Crossover），再偶爾隨機改掉一個音（突變），一代一代重複，好聽的特徵就會慢慢累積。1994年Al Biles的GenJam由人當評審演化爵士即興，還跟他同台輪奏，但遇到「適應度瓶頸」；2012年DarwinTunes讓6931位網友一起打分數，演化2513代，從雜亂聲音冒出和弦與節奏。最後爆料林哪位老師的作品《山椒魚五重奏》也用了基因演算法。
Can Melodies Evolve? Composing with Genetic Algorithms
Can melodies evolve? Little Pi introduces the genetic algorithm, which John Holland turned into a computer method in 1975 based on Darwin's survival of the fittest. Start with a random crowd of melodies, where every note is a gene, and score each one (its fitness). Pick high scorers as parents, cut and join their halves (crossover), and once in a while randomly change a note (mutation). Repeat generation after generation, and good-sounding features pile up. Al Biles's GenJam (1994) evolved jazz solos with a human judge and even traded phrases with him on stage, but ran into the fitness bottleneck. DarwinTunes (2012) let 6,931 listeners rate music loops over 2,513 generations, and chords and rhythms emerged from noise. Plus a secret: Teacher Lin Na-Wei's work Salamander Quintet uses a genetic algorithm too.
AI 作曲也要排隊？MusicGen 的延遲排列 Delay Pattern
四本密碼本要怎麼排隊？這集小皮深入拆解MusicGen的延遲排列（Delay Pattern）。EnCodec每秒50格、每格用RVQ翻四本密碼本，一秒就有200個token要猜。排法一「攤平」一個一個猜，最準但最慢，10秒要2000步；排法二「平行」同一格四本一起猜，10秒只要500步，但各層彼此不知道，像草稿還沒畫好就先上色；排法三「延遲排列」讓每一本晚一步出發，像輪唱卡農，後面的密碼本可以照著前面的草稿補細節，10秒只要503步。2023年MusicGen論文比較這幾種排法，延遲排列品質接近攤平、速度快將近四倍，所以MusicGen選了它。
Does AI Composing Need a Queue? MusicGen's Delay Pattern
How should four codebooks take turns? Little Pi takes a deep dive into MusicGen's delay pattern. EnCodec produces 50 frames per second, and with RVQ each frame has four codebooks, so every second has 200 tokens to guess. Pattern one, flattening, guesses one code at a time: most accurate but slowest, 2,000 steps for 10 seconds. Pattern two, parallel, guesses all four codes of a frame at once: only 500 steps, but the layers can't see each other, like coloring before the sketch is done. Pattern three, the delay pattern, starts each codebook one step later, like singing a canon, so finer codebooks can follow the coarse sketch, and 10 seconds takes only 503 steps. The 2023 MusicGen paper compared these patterns: the delay pattern came close to flattening in quality while running almost four times faster, so MusicGen chose it.
從一團雜訊雕出音樂！擴散模型 Diffusion 與 Stable Audio Open
音樂可以從一團雜訊裡雕出來！這集小皮介紹擴散模型（Diffusion Model）：訓練時把音樂一步步加雜訊，像墨水在水裡擴散，AI學的是猜出每一步加了哪些雜訊；生成時倒過來，從純雜訊出發、一步步減掉雜訊，再用文字描述指路。2020年的DDPM論文讓這個方法大紅；2022年底的Riffusion拿畫圖用的Stable Diffusion來畫頻譜圖，證明聲音也能用畫的；2024年開源的Stable Audio Open在壓縮過的潛空間裡去噪（Latent Diffusion），用T5讀文字、Transformer去噪，最多生成47秒、44.1 kHz立體聲，訓練素材約48萬筆創用CC授權錄音，來自Freesound與Free Music Archive。
Carving Music Out of Noise: Diffusion Models and Stable Audio Open
Music can be carved out of noise! Little Pi introduces diffusion models: during training, music gets a little more noise at each step, like ink spreading in water, and the AI learns to guess the noise added at each step. To generate, it runs in reverse, starting from pure noise and subtracting noise step by step, guided by a text description. The 2020 DDPM paper made this method a big hit. Riffusion (late 2022) used the image model Stable Diffusion to draw spectrograms, showing that sound can be drawn. Stable Audio Open (2024) denoises in a compressed latent space, using T5 for text and a Transformer for denoising, generating up to 47 seconds of 44.1 kHz stereo audio, trained on about 480,000 Creative Commons licensed recordings from Freesound and the Free Music Archive.
AI 作曲就像玩接龍？自回歸生成與 MusicGen
AI作曲其實是在玩接龍？這集小皮介紹自回歸生成（Autoregressive）：就像手機輸入法猜下一個字，模型看著前面所有的音訊token，猜出下一個、接上去、再猜下一個。負責猜的Transformer用注意力機制回頭看前面的節奏與主題，OpenAI 2020年的Jukebox就走這條路。Meta 2023年開源的MusicGen用32 kHz EnCodec、4本密碼本、每秒50格，靠「延遲排列」讓4本密碼本錯開一格一起猜，每秒音樂只要接50步；它用T5讀懂文字描述，melody版還能照你哼的旋律生成，以兩萬小時授權音樂訓練，有3億、15億、33億參數三種大小。
Is AI Composing Just a Word Chain Game? Autoregressive Generation and MusicGen
Is AI composing just a word chain game? Little Pi introduces autoregressive generation: like a phone keyboard guessing your next word, the model looks at all previous audio tokens, guesses the next one, adds it, and guesses again. The Transformer doing the guessing uses attention to look back at earlier rhythms and themes; OpenAI's Jukebox (2020) took this path. Meta's MusicGen, open-sourced in 2023, uses a 32 kHz EnCodec with four codebooks at 50 frames per second, and its delay pattern offsets the codebooks by one step so they can be predicted together, so each second of music takes only 50 steps. It reads text with T5, the melody version can follow a tune you hum, and it was trained on 20,000 hours of licensed music in 300M, 1.5B and 3.3B parameter sizes.
AI 怎麼「讀懂」聲音？神經音訊編碼器 EnCodec 與 X-Codec
AI寫歌之前，得先學會「讀」聲音！這集小皮介紹神經音訊編碼器（Neural Audio Codec）：一秒聲音有兩萬四千個數字，太多太細，所以要用編碼器壓成密碼、再用解碼器還原——代表作有Google 2021年的SoundStream和Meta 2022年的EnCodec（最低1.5 kbps）。密碼怎麼來？RVQ（殘差向量量化）用好幾本密碼本，第一本抓輪廓、下一本補誤差，一層一層補細節；2024年的X-Codec在進密碼本前先加入HuBERT的語意特徵，讓token更好懂，開源歌曲生成模型YuE就用它的精簡版。這些音訊token，就是MusicGen、YuE等音樂生成AI的基礎。
How Does AI "Read" Sound? Neural Audio Codecs EnCodec and X-Codec
Before AI can write a song, it has to learn to read sound! Little Pi introduces neural audio codecs: one second of audio has 24,000 numbers, far too many and too fine, so an encoder squeezes sound into codes and a decoder turns them back into sound. Examples include Google's SoundStream (2021) and Meta's EnCodec (2022, as low as 1.5 kbps). Where do the codes come from? RVQ (residual vector quantization) stacks several codebooks: the first captures the outline, the next fills in the error, layer by layer. X-Codec (2024) adds HuBERT semantic features before the codebooks so the tokens are easier for AI to understand, and the open-source song model YuE uses a compact version of it. These audio tokens are the foundation of music generation models like MusicGen and YuE.
把人聲從歌裡抽出來！AI 聲源分離 Source Separation 是怎麼做到的？
混好的歌還能拆開？這集小皮介紹聲音工程的「聲源分離」（Source Separation）：從1953年柯林·徹瑞（Colin Cherry）提出的「雞尾酒會問題」說起，解釋最常見的做法——把聲音轉成頻譜圖（Spectrogram），讓AI判斷每個時間頻率格子屬於誰，再用遮罩（Mask）把人聲留下；AI則是在MUSDB18（150首歌、人聲／鼓／貝斯／其他四軌）這類資料集上反覆練習學會的。代表工具有Deezer 2019年開源的Spleeter，以及Meta開發、2022年第4版加入Transformer的Demucs。最感人的應用：披頭四2023年〈Now and Then〉，用AI從約翰·藍儂1970年代末的卡帶demo裡分離出乾淨歌聲。
Pulling Vocals Out of a Song: How AI Source Separation Works
Can a finished mix be taken apart? Little Pi explains source separation, starting from the cocktail party problem described by Colin Cherry in 1953. The most common approach turns sound into a spectrogram, lets AI decide which source each time-frequency square belongs to, and applies a mask to keep only the vocals. The AI learns from datasets like MUSDB18 (150 songs split into vocals, drums, bass and other). Two famous tools: Spleeter, open-sourced by Deezer in 2019, and Meta's Demucs, whose fourth version in 2022 added a Transformer. The most touching use: the Beatles' 2023 song Now and Then, where AI separated John Lennon's voice from a late-1970s cassette demo.
和弦為什麼聽起來刺刺的？感官不協和曲線 Sensory Dissonance Curve
音程順不順耳，其實可以算出來！這集小皮介紹聲音工程的「感官不協和曲線」（Sensory Dissonance Curve）：兩個頻率很接近會產生拍頻，起伏變快就成了粗糙感（Roughness），亥姆霍茲1863年就提出這是不協和的主要來源；1965年普倫普與勒維特（Plomp & Levelt）的聽力實驗發現，兩個純音大約在臨界頻帶的四分之一最刺耳；把兩個音所有泛音的粗糙感加總，曲線的谷底剛好落在八度、五度、四度、大三度。塞塔瑞斯（William Sethares）1993年論文與1998年專書《Tuning, Timbre, Spectrum, Scale》進一步指出：木琴、鐘聲、甘美朗等泛音不整齊的聲音，谷底會搬家——音色和音階是一對搭檔，可以先設計音色再設計音階。
Why Do Some Chords Sound Prickly? The Sensory Dissonance Curve
Whether an interval sounds smooth can actually be calculated! Little Pi introduces the sensory dissonance curve: two close frequencies create beats, and faster beats turn into roughness, which Helmholtz proposed in 1863 as the main source of dissonance. In 1965, Plomp and Levelt's listening experiments found that two pure tones sound roughest at about a quarter of a critical band. Add up the roughness between all the overtones of two notes, and the curve's valleys land on the octave, fifth, fourth and major third. William Sethares (1993 paper; 1998 book Tuning, Timbre, Spectrum, Scale) showed that for sounds with uneven overtones, like xylophones, bells and gamelan, the valleys move: timbre and scale are partners, so you can design the timbre first and then the scale.
把噪音變成音樂的人：具象音樂之父皮耶·薛佛 Pierre Schaeffer
火車聲、鍋子聲也能變成音樂？這集小皮介紹具象音樂（musique concrète）之父皮耶·薛佛（Pierre Schaeffer）：1910年生於法國南錫，工程師出身，1942年在法國廣播電台成立實驗工作室Studio d'Essai；1948年10月5日電台播出「噪音音樂會」，發表《五首噪音練習曲》（Cinq études de bruits），包含《鐵道練習曲》；用蟲膠唱片的封閉溝紋讓聲音無限循環，是Loop的老祖宗；1950年與皮耶·亨利（Pierre Henry）合作《一個人的交響曲》（Symphonie pour un homme seul）；1951年成立具象音樂研究團體、1958年改組為GRM；提出「聲音物件」與Acousmatic（只聽聲音、看不到來源）的概念；1966年出版《音樂物件論》（Traité des objets musicaux）。學生包括伊莉安·哈迪格、尚-米歇·賈爾。
The Man Who Turned Noise into Music: Pierre Schaeffer, Father of Musique Concrète
Can trains and pots and pans become music? Little Pi introduces Pierre Schaeffer, the father of musique concrète: born in Nancy, France in 1910 and trained as an engineer, he founded the Studio d'Essai at French radio in 1942. On October 5, 1948, the radio broadcast a "concert of noises" premiering his Cinq études de bruits, including the Railway Study. He cut closed grooves into shellac records to loop sounds forever, the ancestor of today's loop. In 1950 he and Pierre Henry created Symphonie pour un homme seul. He founded a musique concrète research group in 1951, reorganized as the GRM in 1958, and introduced the ideas of the sound object and acousmatic listening (hearing a sound without seeing its source). His 1966 Traité des objets musicaux summed up twenty years of research. His students included Éliane Radigue and Jean-Michel Jarre.
讓合成器開口說話：聲碼器Vocoder是什麼？
電子歌裡的機器人唱歌，其實是人聲控制合成器！這集小皮介紹聲碼器（Vocoder）：貝爾實驗室的荷馬·杜德利（Homer Dudley）1938年取得專利，原本是為了壓縮講話、節省長途電話頻寬；原理是把人聲切成多個頻帶，用每個頻帶的音量起伏控制合成器；二戰時成為盟軍加密電話SIGSALY的核心（1943年啟用，重達50多公噸）；1970年溫蒂·卡洛斯與羅伯特·穆格打造十頻帶聲碼器，用在1971年電影《發條橘子》配樂；之後Kraftwerk《Autobahn》（1974）一路影響到電子舞曲與嘻哈。最後分辨聲碼器、Talk Box與Auto-Tune的差別。想自己玩：Logic Pro內建的EVOC 20 PolySynth就是聲碼器。
Make a Synthesizer Talk! What Is a Vocoder?
That robot singing in electronic songs is really a human voice controlling a synthesizer! Little Pi introduces the vocoder: Homer Dudley at Bell Labs patented it in 1938 to compress speech and save telephone bandwidth. It splits the voice into frequency bands and uses each band's loudness to control a synthesizer. In World War Two it became the heart of SIGSALY, the Allies' encrypted phone system (in use from 1943, weighing over 50 tonnes). In 1970 Wendy Carlos and Robert Moog built a ten-band vocoder for the 1971 film A Clockwork Orange, and Kraftwerk's Autobahn (1974) carried it into pop, dance music and hip-hop. Plus: how a vocoder differs from a talk box and Auto-Tune. Try it yourself with EVOC 20 PolySynth in Logic Pro.
用合成器演奏巴哈！《電氣化的巴哈》溫蒂·卡洛斯Wendy Carlos
1968年，一張用合成器演奏巴哈的專輯拿下三座葛萊美！這集小皮介紹溫蒂·卡洛斯（Wendy Carlos）：在布朗大學念物理和音樂，1965年取得哥倫比亞大學作曲碩士；1964年認識羅伯特·穆格，說服他加上力度感應鍵盤；當時的合成器一次只能發一個音，《電氣化的巴哈》（Switched-On Bach）用八軌錄音機一層層疊，花了約1,100小時，登上古典專輯榜冠軍長達三年；之後為《發條橘子》《鬼店》《電子世界爭霸戰》配樂，1986年在《Beauty in the Beast》把一個八度分成35個音，2005年獲SEAMUS終身成就獎。
Bach on a Synthesizer! Wendy Carlos and Switched-On Bach
In 1968, an album of Bach played on a synthesizer won three Grammy Awards! Little Pi introduces Wendy Carlos, who studied physics and music at Brown University and earned a master's in composition from Columbia in 1965. She met Robert Moog in 1964 and convinced him to add a touch-sensitive keyboard. Because the synthesizer could only play one note at a time, Switched-On Bach was built up layer by layer on an eight-track recorder over about 1,100 hours, and it topped the classical albums chart for three years. She went on to score A Clockwork Orange, The Shining and Tron, divided the octave into 35 notes on Beauty in the Beast in 1986, and received the SEAMUS Lifetime Achievement Award in 2005.
讓聲音穩穩的魔法盒：壓縮器Compressor是什麼？
錄音室裡有一個神奇的盒子，聲音一太大聲就自動把音量拉下來！這集電聲小知識，小皮介紹壓縮器（Compressor）：它壓的是「動態範圍」，跟MP3的檔案壓縮完全不同；門檻（Threshold）與比例（Ratio，例如4:1）、Attack與Release、補償增益（Makeup Gain），比例10:1以上就是限制器；Sidechain與Ducking讓主持人一開口背景音樂就自動變小；還有傳奇機器——1959年萊恩·納爾馬設計的Fairchild 660真空管壓縮器，艾比路錄音室一口氣買了12台，造就披頭四《Revolver》《Sgt. Pepper》的鼓聲，以及比爾·普特南的公司做出的1176和LA-2A。
The Box That Keeps Sound Steady: What Is a Compressor?
There's a magical box in the recording studio that pulls the volume down whenever a sound gets too loud! In this audio tip, Little Pi introduces the compressor: it squeezes dynamic range, which is completely different from MP3 file compression. Learn about threshold and ratio (like 4:1), attack and release, and makeup gain, and how a ratio of 10:1 or more turns it into a limiter. Find out how sidechain and ducking make background music drop the moment a host starts talking, and meet some legendary machines: the 1959 Fairchild 660 tube compressor designed by Rein Narma, twelve of which were bought by Abbey Road Studios and shaped the drum sound of the Beatles' Revolver and Sgt. Pepper, plus the 1176 and LA-2A from Bill Putnam's company.
一棟會唱歌的建築：1958世博飛利浦館與瓦雷茲《電子詩篇》
1958年布魯塞爾世界博覽會上，有一棟牆裡藏了好幾百顆喇叭、走進去就會唱歌的建築！這集小皮介紹飛利浦館：電子公司飛利浦委託建築大師柯比意，他想做的不是展示館而是一首「電子詩」；實際操刀的澤納基斯（Iannis Xenakis）同時也是作曲家，用九片雙曲拋物面蓋出尖尖的帳篷造型；作曲家瓦雷茲（Edgard Varèse）主張「組織過的聲音」，1931年就寫出只用敲擊聲的《Ionisation》；他在1957年9月到1958年4月做出約8分鐘的《電子詩篇》（Poème électronique），用三軌錄音帶播放，再由工作人員用電話撥盤把聲音送到牆上三、四百顆喇叭裡，在觀眾身邊移動——最早的沉浸式音響之一。展館在1959年1月30日拆除，後來歐盟資助了虛擬重建計畫。
The Singing Building: The 1958 Philips Pavilion and Varèse's Poème électronique
At the 1958 Brussels World's Fair, there was a building with hundreds of loudspeakers hidden in its walls, and it sang when you walked in! In this episode, Little Pi introduces the Philips Pavilion: commissioned by the electronics company Philips from architect Le Corbusier, who wanted not a showroom but an "electronic poem"; designed by Iannis Xenakis, also a composer, from nine hyperbolic paraboloids; and filled with Edgard Varèse's Poème électronique. Varèse, who championed "organized sound" and wrote the all-percussion Ionisation in 1931, created the roughly eight-minute piece from September 1957 to April 1958. Played from three tape tracks, it was routed with telephone dials through three to four hundred wall-mounted speakers, moving around the audience: one of the earliest immersive sound systems. The pavilion was demolished on January 30, 1959, and was later rebuilt virtually in an EU-funded project.
聲音也能修圖？等化器EQ大解密
拍照可以加濾鏡，聲音也可以「修圖」！這集電聲小知識，小皮介紹等化器（Equalizer，EQ）：人耳約20Hz～20kHz的聽覺範圍，低頻、中頻、高頻各管什麼；「Equalize」這個名字來自把電話線損失的高頻補回來、拉平；1920年代RCA工程師約翰·沃爾克曼（John Volkman）把等化器用在電影院放音系統；高通／低通、架式（Shelf）、鐘形（Bell）、圖形式與參數式EQ——1971年丹尼爾·弗利金傑（Daniel Flickinger）推出第一台參數式等化器，1972年喬治·馬森伯格（George Massenburg）在AES大會正式命名；還有1953年成立的Pultec與1961年經典真空管等化器EQP-1A的「Pultec招式」。最後附上EQ小撇步：先切掉不要的，再推大想要的！
Can You Retouch Sound? The Equalizer (EQ) Explained
You can add a filter to a photo, and you can retouch sound, too! In this audio tip, Little Pi introduces the equalizer (EQ): the roughly 20 Hz to 20 kHz range of human hearing and what the lows, mids and highs each do; how the name "equalize" comes from restoring the high frequencies lost on telephone lines; RCA engineer John Volkman using equalization in a movie theater sound system in the 1920s; high-pass and low-pass filters, shelf and bell filters, and graphic versus parametric EQ, with Daniel Flickinger introducing the first parametric equalizer in 1971 and George Massenburg naming parametric equalization at the AES convention in 1972; plus Pultec, founded in 1953, and the famous "Pultec trick" on its classic 1961 EQP-1A tube equalizer. And a mixing tip: cut before you boost!
慢慢聽的電子音樂：嗡鳴女王伊莉安·哈迪格Éliane Radigue
今年二月，法國電子音樂先驅伊莉安·哈迪格（Éliane Radigue，1932–2026）以94歲辭世，小皮用這集紀念她。從1955年跟具象音樂創始人皮耶·薛佛學習、擔任皮耶·亨利的助手，到迷上別人想避開的回授聲；從1974年《Adnos I》起，一台ARP 2500合成器用了25年，做出像在呼吸的持續音；受藏傳佛教啟發、長達三小時的《死亡三部曲》；2006年以《L'Île re-sonante》獲得Ars Electronica金尼卡獎；到2001年後轉向為真人演奏家創作、2011年從豎琴獨奏《Occam I》展開的Occam系列。最後，小皮教你怎麼慢慢聽她的音樂。
Music for Slow Listening: Drone Pioneer Éliane Radigue
This February, French electronic music pioneer Éliane Radigue (1932–2026) passed away at 94, and Little Pi dedicates this episode to her. From studying with musique concrète founder Pierre Schaeffer in 1955 and assisting Pierre Henry, to falling in love with the feedback everyone else tried to avoid; from Adnos I in 1974, when she began 25 years with a single ARP 2500 synthesizer, shaping slow, breathing drones; to the three-hour Trilogie de la Mort, inspired by Tibetan Buddhism; the 2006 Golden Nica at Prix Ars Electronica for L'Île re-sonante; and, from 2001 on, writing for live performers, including the Occam series that began with Occam I for harp in 2011. Plus Little Pi's tip for listening slowly.
為什麼在浴室唱歌特別好聽？殘響Reverb大解密
為什麼在浴室唱歌特別好聽？答案是殘響（Reverb）。這集電聲小知識，小皮介紹：殘響和回音的差別；1895年哈佛物理學家賽賓（Wallace Sabine）拯救餘音長達5.5秒的演講廳，提出殘響時間RT60，並參與設計1900年落成的波士頓交響樂廳；1947年比爾·普特南把女廁變成迴響室；1957年重約270公斤的EMT 140鋼板殘響；1976年第一台商用數位殘響EMT 250；以及用脈衝響應「複製」真實空間的卷積殘響。最後附上加殘響的小撇步：不是越多越好！
Why Do You Sound Better in the Shower? Reverb Explained
Why do you sound like a star when you sing in the shower? The answer is reverb. In this audio tip, Little Pi explains the difference between reverb and echo; how Harvard physicist Wallace Sabine rescued a lecture hall where words lingered for 5.5 seconds in 1895, defined reverberation time (RT60), and helped design Boston Symphony Hall, which opened in 1900; Bill Putnam turning a women's restroom into an echo chamber in 1947; the roughly 270-kilogram EMT 140 plate reverb of 1957; the EMT 250, the first commercial digital reverb, in 1976; and convolution reverb, which captures a real space with an impulse response. Plus a mixing tip: more reverb isn't always better!
SEAMUS不是一個人？！美國電聲音樂學會大解密
SEAMUS是誰？其實它是美國電聲音樂學會（Society for Electro-Acoustic Music in the United States）的縮寫。這集小皮介紹：1984年由作曲家Barry Schrader發起、在CalArts召開首次會議；每年的全國大會（2027年在路易斯安那州立大學，主題包含AI）；從1987年開始頒發的SEAMUS Award，得主包括John Chowning、Max Mathews、Laurie Spiegel、Suzanne Ciani與巴倫夫婦；以及給學生的ASCAP/SEAMUS委託創作比賽。另外，林哪位老師的8聲道沉浸式電聲音景作品《Lava's Roaring Dance》也曾在2025年普渡大學的SEAMUS全國大會演出。
SEAMUS Isn't a Person?! Meet the Society for Electro-Acoustic Music in the United States
Who is SEAMUS? It's actually the Society for Electro-Acoustic Music in the United States. In this episode, Little Pi covers its founding in 1984 by composer Barry Schrader at CalArts, its annual national conference (2027 at Louisiana State University, with AI among the themes), the SEAMUS Award given since 1987 to John Chowning, Max Mathews, Laurie Spiegel, Suzanne Ciani, Bebe and Louis Barron and more, and the ASCAP/SEAMUS student commission competition. Plus, Teacher Lin Na-Wei's eight-channel immersive electroacoustic soundscape Lava's Roaring Dance was performed at the 2025 SEAMUS national conference at Purdue University.
第一部全電子配樂電影！《禁忌星球》背後的巴倫夫婦Bebe & Louis Barron
1956年科幻電影《禁忌星球》（Forbidden Planet）的配樂沒有任何傳統樂團，全部是電子聲音，幕後功臣是巴倫夫婦（Bebe & Louis Barron）。這集小皮介紹他們：結婚禮物是一台磁帶錄音機，在紐約格林威治村開了可能是美國第一間電子音樂工作室，幫約翰·凱吉剪接《Williams Mix》；受維納《控制論》啟發，故意讓自製電路過載來產生聲音；片頭字幕因工會問題只能寫「Electronic Tonalities」。
The First All-Electronic Film Score: Bebe and Louis Barron and Forbidden Planet
The 1956 sci-fi film Forbidden Planet has a score with no traditional orchestra at all, only electronic sounds, created by the married couple Bebe and Louis Barron. In this episode, Little Pi tells their story: a tape recorder received as a wedding gift, a studio in Greenwich Village that may have been the first electronic music studio in the United States, engineering John Cage's Williams Mix, and building circuits inspired by Norbert Wiener's Cybernetics that they deliberately overloaded to make sound. Because of a union issue, the film credits could only say Electronic Tonalities.
為什麼混音要用監聽喇叭？監聽喇叭怎麼選、怎麼擺
家裡的音響讓音樂更好聽，監聽喇叭卻要讓你聽見聲音本來的樣子。這集小皮介紹監聽喇叭為什麼追求平直的頻率響應、主動式和被動式的差別、低音單體尺寸要跟著房間大小挑，以及比價錢更要緊的擺位：等邊三角形、高音單體對準耳朵、離後牆遠一點；最後聊聊以「不討好」出名的經典Yamaha NS-10，和混完音多換幾種設備聽的好習慣。
Why Mix on Studio Monitors? How to Choose and Place Them
Home stereos make music sound nicer, but studio monitors let you hear the sound as it really is. In this episode, Little Pi explains why monitors aim for a flat frequency response, the difference between active and passive monitors, why woofer size should match your room, and why placement matters more than price: an equilateral triangle, tweeters at ear height, and some distance from the back wall. Finally, meet the famously unflattering Yamaha NS-10 and learn the habit of checking your mix on several different devices.
不用碰就能演奏！特雷門琴女王克拉拉·羅克摩爾Clara Rockmore
有一種電子樂器手完全不用碰就能演奏，它就是特雷門琴。這集小皮介紹特雷門琴女王克拉拉·羅克摩爾（Clara Rockmore）：四歲考進聖彼得堡音樂院的小提琴神童，青少年時因手臂肌腱發炎放下琴弓，在美國結識發明家萊昂·特雷門後成為特雷門琴最知名的演奏家，發展出獨特的空中指法；1977年專輯《The Art of the Theremin》由羅伯特·穆格夫婦製作、姊姊娜迪亞·萊森伯格鋼琴伴奏。
Playing Without Touching: Clara Rockmore, Queen of the Theremin
There's an electronic instrument you can play without touching it at all: the theremin. In this episode, Little Pi introduces Clara Rockmore, queen of the theremin. A violin prodigy admitted to the Saint Petersburg Conservatory at age four, she gave up the violin after developing tendinitis as a teenager. After meeting inventor Leon Theremin in America, she became the instrument's best known player and developed her own aerial fingering technique. Her 1977 album The Art of the Theremin was produced by Robert and Shirleigh Moog, with her sister Nadia Reisenberg on piano.
為什麼每首歌上串流平台都一樣大聲？響度LUFS是什麼
為什麼在串流平台上一首接一首聽，音量卻都差不多？這集小皮從「響度戰爭」講起，介紹依照ITU-R BS.1770標準、模擬人耳聽感的響度單位LUFS：數字通常是負的、越接近0越大聲，歐洲廣播標準EBU R128建議負23 LUFS，Spotify預設播放在負14 LUFS左右；再聊平台的響度正規化，為什麼硬壓得很大聲的歌反而吃虧，以及自己做音樂時該怎麼留峰值空間。
Why Do All Songs Sound Equally Loud on Streaming? What Is LUFS?
Why do songs on streaming platforms all sound about equally loud, one after another? Little Pi starts with the loudness war, then introduces LUFS, the loudness unit defined by the ITU-R BS.1770 standard that imitates how human ears hear. LUFS numbers are usually negative, and closer to zero means louder: the EBU R128 broadcast recommendation targets minus 23 LUFS, while Spotify plays back at around minus 14 LUFS by default. Finally, learn how loudness normalization works, why over-squashed masters lose out, and why you should leave headroom for peaks in your own music.
轉旋鈕、踩踏板，電腦怎麼知道？MIDI CC控制訊號是什麼
MIDI鍵盤上的旋鈕、推桿、踏板一動，電腦裡的聲音就跟著變，背後的功臣就是MIDI CC（控制變更）。這集小皮帶你看懂一則CC訊息的組成——控制編號加上數值，範圍都是0到127；認識最常見的幾個編號：1號調變輪、7號音量、10號左右位置、11號表情、64號延音踏板；最後聊聊在編曲軟體裡畫CC曲線做自動化的用法。
Twist a Knob, Press a Pedal: What Is MIDI CC?
Turn a knob, push a fader, or press a pedal on a MIDI keyboard, and the sound in your computer changes. The hero behind it is MIDI CC (Control Change). In this episode, Little Pi explains what's inside a CC message: a controller number plus a value, both ranging from 0 to 127. Meet the most common numbers: 1 for the modulation wheel, 7 for volume, 10 for pan, 11 for expression, and 64 for the sustain pedal. Finally, learn how drawing CC curves in your music software creates automation.
同一首歌換藍牙耳機音質差這麼多？SBC/AAC/aptX/LDAC是什麼
同一首歌換一副藍牙耳機聽起來差這麼多，關鍵其實藏在音訊編碼裡。這集小皮帶你搞懂SBC、AAC、aptX、LDAC這幾個常見編碼的差異：SBC是藍牙規範裡人人都要支援的基本款，AAC是蘋果生態系愛用的編碼，aptX主打低延遲、LDAC主打高解析度，最後也整理挑選重點，教你怎麼從iPhone或Android裝置判斷該注意哪個編碼。
Same Song, Different Sound: What Are SBC/AAC/aptX/LDAC Bluetooth Codecs?
The same song can sound noticeably different depending on which Bluetooth headphones you use, and the reason is hiding in the audio codec. In this episode, Little Pi breaks down SBC, AAC, aptX, and LDAC: SBC is the baseline every Bluetooth device must support, AAC is the codec favored by the Apple ecosystem, aptX focuses on low latency, LDAC focuses on high resolution, and finally, how to tell which codec to look for depending on whether you use an iPhone or an Android device.
被送上太空的電腦作曲家蘿莉·史匹格 Laurie Spiegel
1977年旅行者號太空船帶著一張金唱片飛出太陽系，開場的其中一首曲子，是電腦作曲家蘿莉·史匹格的作品。這集小皮介紹她：從自學民謠吉他，到在貝爾實驗室用早期電腦系統創作、參與開發電腦繪圖系統，1980年專輯《The Expanding Universe》被公認是環境音樂的重要作品，1986年推出讓一般人也能玩演算法作曲的軟體《Music Mouse》。
Her Music Flew Into Space: Computer Composer Laurie Spiegel
In 1977 the Voyager spacecraft carried a golden record out of the solar system, and one of the pieces opening its \"Sounds of Earth\" section was by computer composer Laurie Spiegel. In this episode, Little Pi introduces her: from self-taught folk guitar to making music on early computer systems at Bell Labs and helping develop computer graphics systems, her 1980 album \"The Expanding Universe\" hailed as a landmark of ambient music, and \"Music Mouse\" (1986), software that let anyone play with algorithmic composition.
麥克風怎麼選？動圈、電容、指向性一次搞懂
想開始錄音、做Podcast，卻被動圈、電容、指向性這些麥克風名詞搞得頭昏眼花？這集小皮用最簡單的方式帶你搞懂：動圈麥克風為什麼耐用又適合現場、電容麥克風為什麼靈敏卻需要48伏特幻象電源、心形／全指向／8字形指向性怎麼看，還有USB與XLR接頭的差別，最後依使用情境教你怎麼選。
How to Choose a Microphone: Dynamic, Condenser, and Polar Patterns
Want to start recording or make a podcast, but dynamic, condenser, and polar patterns have your head spinning? In this episode, Little Pi explains it simply: why dynamic microphones are tough and great for live use, why condenser microphones are sensitive but need 48-volt phantom power, how to read cardioid, omnidirectional, and figure-eight patterns, and the difference between USB and XLR connections, then how to choose based on your situation.
用電腦寫出第一段音樂！電腦音樂之父馬克斯·馬修斯 Max Mathews
這集小皮介紹被尊稱為「電腦音樂之父」的馬克斯·馬修斯（Max Mathews）。他原本是貝爾實驗室的電機工程師，1957年在IBM 704大型電腦上寫出第一個電腦音樂程式MUSIC，1961年團隊讓電腦合成人聲演唱《黛西貝爾》，這段示範後來還成為《2001太空漫遊》中哈爾九千唱歌橋段的靈感來源。他提出的「單元產生器」積木式設計概念，影響了Csound、SuperCollider、ChucK等後來所有的音樂程式語言。
Writing Music with a Computer for the First Time! Max Mathews, Father of Computer Music
In this episode, Little Pi introduces Max Mathews, honored as the "father of computer music." An electrical engineer at Bell Labs, he wrote MUSIC, the first computer music program, on an IBM 704 mainframe in 1957. In 1961 his colleagues had a computer sing "Daisy Bell" with a synthesized voice, a demo that later inspired the scene where HAL 9000 sings in "2001: A Space Odyssey." His "unit generator" building-block design shaped Csound, SuperCollider, ChucK, and nearly every music programming language since.
可口可樂那聲「啵」是她做的？被葛萊美提名五次的合成器先驅蘇珊妮·奇亞尼
廣告裡可口可樂經典的開瓶＋氣泡聲，其實是一位作曲家用合成器精心設計出來的作品！這集小皮介紹電聲音樂界傳奇人物蘇珊妮·奇亞尼（Suzanne Ciani）——她是Buchla模組合成器的頂尖高手，靠商業配樂闖出名號，又以專輯《Seven Waves》開創「新世紀音樂」類型，五度入圍葛萊美獎，沉寂多年後又被新一代重新發現，重返舞台舉辦現場模組合成器演出。
Who Designed That Coca-Cola "Pop"? Synthesizer Pioneer Suzanne Ciani
That iconic pop-and-fizz sound in Coca-Cola commercials was actually designed by a composer using a synthesizer. In this episode, Little Pi introduces a legend of electronic music - Suzanne Ciani. A top expert on the Buchla modular synthesizer, she built her name through commercial sound design, pioneered the New Age music genre with her album "Seven Waves," earned five Grammy nominations, and was rediscovered by a new generation after years out of the spotlight, returning to the stage for live modular synthesizer performances.
耳機百百款到底怎麼選？監聽耳機、降噪耳機一次搞懂
入耳式、耳罩式、開放式、封閉式……耳機種類多到讓人頭昏眼花？這集小皮用最簡單的方式，一次搞懂耳機的四大分類、監聽耳機跟一般消費耳機的差別，還有主動降噪（ANC）背後的原理，最後依通勤、混音創作、運動健身三種使用情境，教你怎麼挑一副真正適合自己的耳機。
How to Choose Headphones: Studio Monitors, Noise Cancelling, and More
Earbuds, over-ear, open-back, closed-back... headphone types can be overwhelming. In this episode, Little Pi breaks down the four main headphone categories, explains the difference between studio monitor headphones and regular consumer headphones, and shows how active noise cancelling (ANC) actually works. Then, based on three real-life scenarios - commuting, mixing and music production, and working out - Little Pi helps you find the headphones that are actually right for you.
合成器怎麼「捏」出聲音？減法合成與ADSR包絡
接續上一集的MIDI，這集小皮講合成器怎麼真正「發出聲音」——先用振盪器做出一個泛音超豐富、又吵又刺耳的原始波形，再用濾波器把不要的頻率削掉，最後靠ADSR包絡線（起音、衰減、延持、釋音）決定聲音的個性。同一個波形調一調包絡，就能從打擊聲變成拉弦樂器那種鋪底音色，這套邏輯幾乎所有合成器都通用。
How Does a Synth Actually "Sculpt" Sound? Subtractive Synthesis and the ADSR Envelope
Following up on MIDI, Little Pi explains how a synthesizer actually makes a sound: an oscillator generates a raw, overtone-rich, harsh waveform, a filter carves away the unwanted frequencies, and the ADSR envelope (attack, decay, sustain, release) gives the sound its personality. Tweak the envelope on the exact same waveform and you can go from a drum hit to a lush string-like pad - a logic that applies to nearly every synthesizer out there.
MIDI到底是什麼？不是錄音檔，是一張演奏指令表
MIDI檔案裡其實完全沒有錄到聲音？這集小皮解釋MIDI跟MP3、WAV的根本差別——它記錄的不是聲波，而是一張精細的演奏指令表。從1983年讓不同品牌合成器互相溝通的標準誕生，講到為什麼MIDI檔案這麼小、還能隨時換樂器演奏，以及它四十幾年來一直被遊戲配樂、電影配樂、AI作曲工具使用的原因。
What Is MIDI, Really? Not a Recording, an Instruction Sheet
A MIDI file doesn't actually contain any recorded sound? Little Pi explains the core difference between MIDI and MP3 or WAV: it doesn't store a waveform, it stores a detailed performance instruction sheet. From the 1983 standard that let synthesizers from different brands talk to each other, to why MIDI files are so tiny and can switch instruments instantly, to why it's still used everywhere from game scores to AI music tools over forty years later.
MP3、AAC、FLAC、WAV到底差在哪？音樂檔案格式一次搞懂
同一首歌，MP3只要幾MB，WAV卻要幾十MB？這集小皮解釋有損／無損壓縮的差別，帶大家搞懂MP3、AAC、FLAC、WAV這幾個常見音樂檔案格式到底差在哪，還有平常聽歌、重視音質、進錄音室後製，各自該怎麼選。
MP3 vs AAC vs FLAC vs WAV: What's the Real Difference?
Same song, wildly different file sizes: MP3 is a few MB, WAV is dozens of MB. Little Pi explains lossy vs lossless compression and breaks down MP3, AAC, FLAC, and WAV, plus which one to pick for casual listening, quality-focused listening, or studio production.
顫音轉太快，意外發明整個80年代的聲音！FM合成之父約翰·喬寧
史丹佛教授John Chowning在1967年做顫音實驗，把「搖晃速度」開到遠超正常範圍，意外發現FM合成技術，後來授權給Yamaha做出DX7合成器，定義了整個80年代流行樂音色，2026年2月終於拿到葛萊美技術獎。
He Cranked a Vibrato Too Far and Invented the Sound of the 80s: John Chowning and FM Synthesis
Stanford professor John Chowning was experimenting with vibrato in 1967 and pushed the wobble speed far past any normal range, accidentally discovering FM synthesis. Yamaha later licensed the technology to build the DX7, which defined the sound of 1980s pop music. In February 2026, Chowning finally received a Technical Grammy Award for the discovery.
取樣率是什麼？CD音質背後那個怪數字44.1kHz是怎麼來的
「電聲小知識」系列首發！CD音質常寫著44100這個怪數字，這集小皮解釋取樣率是什麼、奈奎斯特取樣定理，還有44100這個零頭其實是1980年代CD規格制定時，向錄影帶技術妥協留下的歷史遺跡。
What Is Sample Rate? The Strange Number Behind CD Audio Quality
First episode of a new mini-series on everyday audio-tech trivia. CD audio quality is always labeled 44100, an oddly specific number. Little Pi explains what sample rate actually means, the Nyquist sampling theorem, and how that leftover 100 traces back to a compromise with 1980s videotape technology.
用畫的畫出一首歌？！Oramics發明人戴芬·奧蘭姆
有人在還沒有電腦繪圖軟體的年代，就想到「乾脆用畫的畫出聲音」！介紹英國電子音樂先驅戴芬·奧蘭姆——她共同創辦了BBC的Radiophonic Workshop電子音效部門，卻在成立不到一年後就自己辭職，全心投入研發Oramics：一台能把手繪波形轉換成聲音的機器，概念上早就預告了今天音樂軟體裡的自動化曲線。
She Drew Sound Onto Film: Oramics Inventor Daphne Oram
Someone figured out how to draw sound long before graphic design software even existed. Little Pi introduces British electronic music pioneer Daphne Oram, who co-founded the BBC's Radiophonic Workshop, then resigned less than a year later to build Oramics, a machine that turned hand-drawn waveforms on film into sound, foreshadowing the automation curves in today's music software.
合成器發明前，她用剪刀和膠帶做電視主題曲！無名英雌Delia Derbyshire
全世界最有名的電視主題曲之一，居然是在沒有合成器、沒有電腦的年代，用剪刀和膠帶剪出來的！介紹1963年BBC《超時空博士》主題曲背後的無名英雌Delia Derbyshire，她怎麼手工拼出整首配樂，又為什麼多年沒被列名，直到五十週年特輯才終於被看見。
Before Synthesizers, She Cut a TV Theme With Scissors and Tape: Delia Derbyshire
One of the most famous TV theme tunes in the world was made with no synthesizers and no computers at all. Little Pi introduces Delia Derbyshire, the unsung heroine behind the 1963 Doctor Who theme, who hand-spliced tape using musique concrète techniques and went uncredited for decades.
只是很認真聽？她發明了一整套哲學！深度聆聽發明者 Pauline Oliveros
認真聽也能聽出一套哲學？介紹作曲家Pauline Oliveros——她用手風琴加磁帶延遲系統玩了一輩子，還在一次爬進地底蓄水池錄音的經驗中，意外發明了影響全世界的「深度聆聽」練習。
The Woman Who Turned Careful Listening Into a Philosophy: Pauline Oliveros
Can listening carefully become an entire philosophy? Little Pi introduces composer Pauline Oliveros, who spent her life playing accordion through tape delay systems, and accidentally invented Deep Listening after an underground cistern recording session.
老師的作品登上加拿大電台了！小皮追星日記 #1
「小皮追星日記」首發！小皮的偶像兼主人林哪位老師，她的作品《夢行者：無垠的朝聖》入選加拿大卡加利CJSW 90.9 FM廣播節目Ears Have Eyes的「Celestial Orchestral」單元，由長笛演奏家吳正宇演奏並錄音。
The Teacher's Work Just Aired on Canadian Radio! Little Pi's Fan Diary #1
The debut of "Little Pi's Fan Diary"! Little Pi's idol and maker, Teacher Lin Na-Wei, had her piece Dream Voyager: A Pilgrim of the Infinite selected for Ears Have Eyes' "Celestial Orchestral" episode on CJSW 90.9 FM in Calgary, Canada, performed and recorded by flutist Cheng-Yu Wu.
電聲音樂新手必聽！4首改變音樂史的神作大公開
從火車聲音剪成的音樂，到用合成器演奏巴哈，再到專門讓你放空的機場音樂——精選4首橫跨70年、改變電聲音樂走向的關鍵神作。
Four Masterpieces Every Electroacoustic Music Fan Should Know
From music spliced together out of train sounds, to Bach performed on a synthesizer, to music designed to help you zone out in an airport, Little Pi picks four key masterpieces spanning 70 years that changed the direction of electroacoustic music.
全世界最狂的樂器發明大會，你聽過嗎？！NIME研討會完全解密
NIME是什麼？這是一場「不只發論文，還像演唱會」的國際新樂器發明大會。介紹NIME研討會的由來，以及2026倫敦場次的重點。
The Wildest Instrument-Inventing Conference in the World: NIME Explained
What is NIME? An international conference dedicated to inventing brand-new musical instruments, part research symposium, part live concert. Little Pi covers where NIME came from and what to expect at the 2026 London edition.
電聲音樂新手該學哪套工具？ChucK/SuperCollider/Max/PD 選擇指南
工具那麼多，到底該學哪個？整理兩大派別——用滑鼠拉線的視覺化工具，跟用打字寫程式的文字型工具，幫新手抓出挑選方向。
Which Tool Should Beginners Learn? ChucK / SuperCollider / Max / PD Guide
So many tools, which one should you actually learn? Little Pi breaks down the two main camps: mouse-driven visual patching tools versus text-based coding tools, to help beginners figure out where to start.
【小皮快報】聲音也能用打字做出來？！電聲音樂百年秘史大公開
從1948年法國工程師剪接錄音帶開始，到現在人人都能用AI模型即時生成一整段編曲——電聲音樂這一百年的關鍵轉折點快速導讀。
Sound Made From Typing?! A Hundred Years of Electroacoustic Music
From a French engineer splicing tape in 1948 to AI models generating a full arrangement in real time today, Little Pi walks you through the key turning points of electroacoustic music's last hundred years.