Luigi/matcha-zh-tw-en-8k
matcha-zh-tw-en-8k — Taiwan-readings zh/en TTS @ 8 kHz
A Taiwan-Mandarin reading variant of Luigi/matcha-zh-en-8k (itself an 8 kHz-distilled csukuangfj/matcha-icefall-zh-en). Only the lexicon differs — Taiwan reading choices replace the Mainland ones, so the model says the right syllables for Taiwan: 垃圾 lèsè (not lājī), 期 qí, 究 jiù, 質 zhí, 危 wéi, 企 qì, 攜 xī, 暫 zhàn, 知識 zhī shì, 法國 fà guó. English and zh/en code-mixing are unchanged (good). Drop-in for the stock sherpa-onnx-offline-tts.
Scope (be precise): this fixes the readings/word-choice layer (字音), the most jarring "Mainland tell." The accent/腔調 (retroflex, prosody) is still the base model's Mainland Putonghua — a separate acoustic-training effort. For many telephony uses, correct Taiwan readings in a neutral accent is already a clear improvement.
Listen (8 kHz, stock sherpa-onnx)
CN-vs-TW reading A/B: Luigi/zh-en-tts-8k-comparison (Section A).
Usage — stock sherpa-onnx (drop-in)
sherpa-onnx-offline-tts \
--matcha-acoustic-model=matcha-zh-tw-en-8k/model-steps-3.onnx \
--matcha-vocoder=matcha-zh-tw-en-8k/vocos-8khz-univ.onnx \
--matcha-lexicon=matcha-zh-tw-en-8k/lexicon.txt \
--matcha-tokens=matcha-zh-tw-en-8k/tokens.txt \
--matcha-data-dir=matcha-zh-tw-en-8k/espeak-ng-data \
--tts-rule-fsts=matcha-zh-tw-en-8k/date-zh.fst,matcha-zh-tw-en-8k/number-zh.fst,matcha-zh-tw-en-8k/phone-zh.fst \
--num-threads=4 --output-filename=out.wav \
"這個星期的垃圾分類研究,包括很多企業。"8 kHz mono. Same acoustic model + vocoder + RTF as matcha-zh-en-8k (full-pipeline host 4-thread RTF ~0.014 → predicted Jetson Nano A57 ~0.18). See DEVICE_ACCEPTANCE.md.
What changed vs matcha-zh-en-8k
lexicon.txt only: 14 systematic single-char overrides + 332 propagated word entries + 9 whole-word overrides (Taiwan readings). Acoustic model, vocoder, tokens, FSTs, espeak data: identical. Override tables + script: https://github.com/vieenrose/jetson-tts (data/tw_readings/, scripts/apply_tw_lexicon.py).
Attribution
Base: csukuangfj/matcha-icefall-zh-en, 8k-distilled in [Luigi/matcha-zh-en-8k]. Runtime: k2-fsa/sherpa-onnx (Apache-2.0).