feat(kokoro): train and verify bilingual PyTorch weights

[Alex-Wengg]

Review the model

Model card · Reproduction commands

Adds a working PyTorch pipeline for adapting Kokoro v1.1-zh to real EMIME MF5 English/Mandarin recordings: acquisition, preprocessing, training, restart, development-only checkpoint selection, export, and offline inference. The selected model completed 500 deterministic updates using 225 available training utterances.

Includes a current model card and compact results with artifact hashes. Weights, real recordings, and generated evaluation audio remain local. Core ML is deferred. This PR is stacked on #94.

Measured quality

The selected checkpoint was frozen before evaluating 74 held-out recordings.

Metric	Baseline	Trained
English WER, 43 utterances	6.15%	4.31%
Mandarin CER, 31 utterances	1.50%	1.93%

Mixed-control raw CER increases from 13.18% to 19.38%; the recognizer sometimes translates mixed speech. This is an experimental model: bilingual listening/tone acceptance, real code-switch training coverage, session-disjoint evaluation, and production voice consent remain unresolved.

Validation

• Recorded model checks: 688 baseline state tensors and 15 parity cases; real-data gradients and micro-overfit; bit-exact restart; 12 exported-model parity cases; offline CPU/GPU synthesis.
• 53 tests, Ruff, and dependency-lock consistency passed before the audio/documentation cleanup. Updated documentation links and Git whitespace checks pass; existing demo assets are unchanged.

Full historical run reports remain linked from the model card.