corpus — phoneme target list + recording protocol #

the audio side of the jeffrey harness. ~40 short jeffrey-pvc clips that the fitter optimizes pink trombone parameters against. each clip is the same phonetic target across multiple carrier contexts so co-articulation effects are observable, not a confound.

scope #

we are not trying to clone arbitrary jeffrey speech. the harness's job is to land coherent vowels and continuants — the things PT models well — and to handle plosives "good enough" for AC's pixelated-voice aesthetic (per reference_pink_trombone.md). the corpus reflects that weighting.

(numbers approximate; final list lives in corpus/phonemes.json for code.)

carrier contexts #

each phoneme is recorded in three contexts so the fitter sees it under different articulatory loads:

1. isolation — sustained for ~1.5s. e.g. for /iː/: just "eeeee". target a steady-state tract pose.
2. digit carrier — embedded inside a counted "one… two… three…" or "a… b… c…" sequence. inherits jeffrey-pvc's prosodic envelope.
3. CVC word — a real english word containing the phoneme in stressed position. provides a co-articulation reference.

three contexts × ~37 phonemes = ~111 prompts, ~1.5s each → ~170s of audio.

we keep the carrier sentence list short (a b c, 1 2 3, "buy", "see", etc.) on purpose. ElevenLabs charges per character, and the per-prompt costs add up if we get expansive. memory: feedback_jeffrey_pvc_settings.md — stability ≥ 0.5 to keep voice identity.

recording settings #

• voice: provider=jeffrey, voice=neutral:0 (the existing jeffrey-pvc PVC)
• stability: 0.5 (memory: identity drift below 0.5)
• similarity: 0.9 (memory)
• style: 0.0 (we want neutral articulation, not performance)
• speed: 0.85 for sustained vowels (gives the fitter ~20% more steady-state to chew on); 1.0 for digit/CVC carriers
• format: server returns mp3; the recorder transcodes to mono 22050 Hz 16-bit PCM WAV for librosa/MFCC compatibility (PT's internal sample rate is 48k but the fitter doesn't care, only the spectral envelope matters)
• with-timestamps: on — char-level alignment means the fitter can crop directly to the target phoneme's onset/offset without re-running whisper

cache key: sha256 of { text, voice, stability, similarity, style, speed } — same convention as pop/bin/say.mjs. reruns of record-corpus.mjs are free unless the prompt list changes.

storage layout #

corpus/
  phonemes.json                       — master list (committed)
  raw/                                — gitignored
    iso/<phoneme-id>.{wav,json}       — iso recordings + alignment sidecar
    digit/<phoneme-id>.{wav,json}     — digit-carrier recordings
    cvc/<phoneme-id>.{wav,json}       — CVC-word recordings
  cropped/                            — gitignored
    <phoneme-id>__<context>.wav        — cropped to target phoneme span
                                        using alignment timestamps
  manifest.json                        — committed; phoneme → file map +
                                        recording metadata + sha hashes

raw/ and cropped/ are gitignored because (a) they're regenerable, and (b) the WAVs are jeffrey-pvc derivatives — the same closed-clone reasoning that keeps references/ out of the repo, applied conservatively.

what we do not record #

• spontaneous jeffrey speech. the goal is to fit a tract, not to capture jeffrey's natural prosody. natural prosody comes back at synth time, not at fit time.
• entire words at fitting resolution. PT can't synthesize an arbitrary english word coherently from a static parameter set; that's render.mjs's job, not the corpus's.
• music or singing. if singing voice ever becomes a target, that's a separate corpus with --style 0.6+ and per-pitch sustains.

next #

• write corpus/phonemes.json
• implement bin/record-corpus.mjs (mirrors pop/bin/say.mjs cache pattern)
• verify: one isolation /iː/ recording, listened to, sounds like jeffrey
• generate full corpus
• crop to per-phoneme spans using alignment sidecars