Luigi/x-asr-zh-tw-en-streaming-ft75m

X-ASR zh-TW/en — fine-tuned 75.5M streaming (native Traditional)

A 75.5M bilingual (Chinese + English) streaming zipformer transducer, int8 ONNX for sherpa-onnx, emitting Traditional Chinese directly (s2twp-relabeled tokenizer, no runtime OpenCC).

• —Warm-started from `pfluo/k2fsa-zipformer-chinese-english-mixed` (75.5M).
• —Fine-tuned on ~1,560 h of Taiwan-Mandarin speech (Breeze-ASR-25 pseudo-labels) + Common Voice zh-TW + NTU-ML.
• —CV-zh-TW (500 clips): recognition CER 0.0822 (warm baseline 0.0845; 30.9M small model 0.108).
• —Speed: int8 RTF 0.032 @ 2 threads — faster than the deployed 480 ms X-ASR (0.037), so it runs real-time on a Jetson Nano gen1 (2 CPU cores). Bilingual zh-en code-switch supported.

Files: encoder.int8.onnx, decoder.onnx, joiner.int8.onnx, tokens.txt (Traditional).

Provenance / license

Research model. Fine-tuning data includes YouTube audio used for recognition training only (the model cannot reproduce the source audio). Not cleared for commercial deployment without retraining on a rights-clean corpus. Recipe + benchmarks: github.com/vieenrose/jetson-stt.