VOICE_V11 acceptance — natural voice: softer onset, breathing pace, integrated breath

Status: Accepted for VOICE_V11. The --version string reports dootdoot VOICE_V11. This note records the four tuning slices behind the freeze; the binding contract is the regenerated golden WAV fixtures (dootdoot-core/tests/fixtures/golden/*.wav, compared byte-for-byte), not the prose here. Acceptance was by-ear across a round of feedback.

Source

Direct listening feedback over several iterations:

1. a plain multi-word phrase ("Hello my name is Sandile") sounded percussive and staccato — onsets too hard-edged;
2. it wanted more change of speed throughout, without leaning on punctuation;
3. a dash ("Hello my name is Sandile - I love cake!!") put a strange amount of noise on the first sentence; and
4. the breathiness sounded artifacty — a separate hiss rather than breath.

The four slices (each its own concern)

• T-109 — soften the attack. Envelope attack ramp 6 ms → 15 ms (quadratic ease-in unchanged, so onsets bloom); the per-word onset transient is quieter (ATTACK_TRANSIENT_MIX 0.07 → 0.04) and longer (ATTACK_TRANSIENT_SECONDS 20 ms → 30 ms) — a breathy consonant, not a pluck.
• T-110 — breathing pace. syllable_rubato_scale(index, total, emphasized) gives each syllable a deterministic duration multiplier: a sinusoidal lilt (±8%, ~5.7-syllable period), agogic lengthening on emphasized syllables (+10%), and phrase-final lengthening (+10%). Gated on the explicit text path; a single-syllable phrase returns exactly 1.0.
• T-112 — localize the dash breath. A dash used to tag its whole preceding clause as a breathy Hesitation (a 0.45 roughness floor on every syllable). Now only the syllable carrying the dash is the hesitation; the preceding clause reads as a plain ChattyReply statement that trails off (not an inquisitive Probe). A single-word filler before a dash still reads as a hesitation.
• T-113 — integrate the breath. The breath was stationary value-noise (a buzzy ~6.3 kHz comb from its fixed 7-sample stride) cross-faded in, so it read as a separate hiss — the classic artifact (Klatt; breathy-vowel synthesis). It is now pitch-synchronously amplitude-modulated (peaking at the glottal closure instant, breath_closure_modulation), sourced from a near-white spectrum shaped by the formant filter, and mixed additively over the tone.

Method & research

The breath fix is grounded in formant-synthesis literature: stationary noise is "perceived as coming from a separate sound source," solved by giving the noise "a temporal envelope of the same periodicity as the pulse train," with more power near the glottal closure instant, mixed additively (Klatt synthesizer; "Synthesis of breathy vowels"; Stanford CCRMA glottal-excitation modeling). By-ear review on the feedback phrases plus the standing corpus (scripts/bb8-metrics) confirmed the onset reads softer, pace varies within an unpunctuated phrase, the first sentence before a dash is no longer a wall of breath, and the breath fuses into the voice — while the droid character is unchanged.

Determinism & contract

• All changes stay inside the fixed, deterministic, bounded droid parameter space (NFR-16); the rubato and the breath modulation are closed-form functions (syllable index / glottal phase) — no randomness was added.
• The rubato (T-110) is gated on the explicit text path and returns 1.0 for a single-syllable phrase, so it does not move the hand-built / empty-chirp / neutral-curve path on its own. The attack (T-109) and breath (T-113) are global synthesis changes that shift every render — which is what the version bump is for; no path is byte-identical to VOICE_V10.
• Sample counts are unchanged on every path (the attack/breath reshape amplitude within fixed-duration syllables; the rubato preserves hand-built durations). The utterance-estimate-vs-render and double-run determinism tests confirm both.
• ACTIVE_VOICE is VOICE_V11; the golden WAV fixtures were regenerated and the double-run determinism test passes.