3-pass diagnostic: naturalize utterances + add per-model raw transcripts vs ground truth (errors highlighted); correct the P2 English OOD finding

Add 3-pass phonemic diagnostic (zh 38/38, en 39/39, entities) + teacher/v1/v2 benchmark (X-ASR+Breeze) + diagnosis

Add pronunciation diagnostic: X-ASR vs Breeze-ASR-25 cross-check (teacher/v1-4.63M/v2-6.85M) on the entity sentence

README: full v4 architecture/training/dataset + reproduction; v3 kept in v3_4.6M/

Keep best 4.63M (v3-30k) ONNX in v3_4.6M/ for record/rollback

README: v4-50k now default (6.85M, clarity+timing); v3 kept for rollback

add long-text chunking synth

export: disable non-ONNX group planner

v4-50k acoustic checkpoint

Default -> v4-50k (6.85M: +contextual-predictors, mrstft clarity, from-scratch); better timing+clarity by ear vs v3

card: v3 pronunciation fix (ㄜ/ㄟ/ㄡ + ㄭ), zh-CER .106->.087

symbol_table: add ㄭ (id 87) for empty-rime vowel

frontend: emit ㄭ for empty-rime syllables

v3 acoustic ckpt 30k (pronunciation fix)

v3 pronunciation fix (ㄜ/ㄟ/ㄡ+ㄭ): acoustic_decoder.onnx

v3 pronunciation fix (ㄜ/ㄟ/ㄡ+ㄭ): acoustic_encoder.onnx

Card: ship 60k (clarity PESQ 3.22 + zh-CER .106), update metrics

Clarity: 60k acoustic checkpoint

Clarity: ship 60k acoustic (sharper mel) — acoustic_decoder.onnx

Clarity: ship 60k acoustic (sharper mel) — acoustic_encoder.onnx

Model card: 24kHz CC0 release, entity-coverage corpus (29.5k), updated metrics

CC0 24kHz vocoder final (snake_v2mid)

CC0 24kHz acoustic ckpt 35k

CC0 24kHz model: meta.json (acoustic 35k + vocoder final)

CC0 24kHz model: vocoder.onnx (acoustic 35k + vocoder final)

CC0 24kHz model: acoustic_decoder.onnx (acoustic 35k + vocoder final)

CC0 24kHz model: acoustic_encoder.onnx (acoustic 35k + vocoder final)

text_norm: time-of-day + decimals

README: reference voice = Mozilla Common Voice zh-TW (CC0/public-domain, commercial-clear), replacing Edge-TTS

gen: pre-normalize text via text_norm (correct entity reading; teacher audio matches manifest)

README: one-command train-on-your-own-voice section

one-click rebuild_voice.sh + generators + text pools

one-click rebuild_voice.sh + generators + text pools

one-click rebuild_voice.sh + generators + text pools

one-click rebuild_voice.sh + generators + text pools

one-click rebuild_voice.sh + generators + text pools

one-click rebuild_voice.sh + generators + text pools

one-click rebuild_voice.sh + generators + text pools

one-click rebuild_voice.sh + generators + text pools

fix: aligner uses text_norm.normalize (entity normalizer)

frontend: entity-aware normalizer

frontend: entity-aware normalizer

remove superseded BreezyVoice-era script

remove superseded BreezyVoice-era script

remove superseded BreezyVoice-era script

remove superseded BreezyVoice-era script

reproduction: actual VoxCPM2-TW pipeline scripts + master run + eval set

reproduction: actual VoxCPM2-TW pipeline scripts + master run + eval set

reproduction: actual VoxCPM2-TW pipeline scripts + master run + eval set

reproduction: actual VoxCPM2-TW pipeline scripts + master run + eval set