Accent profiles and singing styles

Goal

Support selectable accents and singing styles, similar in spirit to Vocaloid's per-language voicebanks and singing styles, but done with rules and parameters rather than new recordings. Because CHIMERA produces sound from articulation, an accent is mostly data the engine already uses: pronunciation rules, vowel targets, timing and consonant details.

Motivation

The current English voice has an unintended accent that some listeners hear as Indian English, and it's still awkward on some words and phrases. Making accent explicit would fix the default voice and add new voices at the same time.

Suspected causes of the current accent (not yet checked by ear)

• phonetics/syllables.rs keeps full vowels in sung function words on purpose ("the" and "a" never shrink to a weak "uh"). That gives a syllable-timed rhythm, one of the strongest cues of Indian English.
• Aspiration after /p t k/ (planner.rs, durations.rs) may be too weak in practice. Unaspirated stops are typical of Indian English.
• Tongue-tip targets for /t d/ (and /n l/) may sit far enough back to sound slightly retroflex.
• Diphthongs like the vowels in "day" and "go" may sound like single held vowels.

Proposed scope

Accent profiles

An accent setting that makes the current hard-coded American rules pluggable:

• Pronunciation rules (phonetics/english.rs): r-dropping after vowels, "bath" with the "father" vowel, vowel mergers, where /t/ is tapped or replaced by a glottal stop, dark vs. clear /l/.
• Vowel targets (vowelTargets.generated.json): a table per accent, fitted to published formant data with the same process that produced the current table.
• Timing and intonation (durations.rs, planner.rs): stress-timed vs. syllable-timed rhythm, how much unstressed vowels shrink, pitch accent shape.
• Consonant details: how much breath follows /p t k/, where the tongue touches for coronals (/t d n l/).

First two accents: General American (fix the current voice) and RP, the standard southern British accent.

Singing styles

Presets for the voice source and singing rules: vibrato rate and depth, breathiness, vocal fry, scoops and falls into notes, larynx height, twang (narrowed bright throat), belt vs. head voice, and vowel shifts on high notes. Most of these are existing glottis and anatomy parameters.

Notes

• The hard part is tuning each accent by ear, not the architecture.
• The compare / hill-climb tooling could help score accent fidelity against reference recordings.