Luigi/moss-transcribe-diarize-zhtw-onnx

Note (2026-07): the ONNX graphs here are the v5 lineage (current for ONNX/sherpa runtimes). The actively developed lineage is v7 (streaming 45 s audio-KV + sentence-cadence markers + CE-only int4 QAT for q4-robustness; q4-GGUF ASCEND all 0.317), shipped as GGUF for RapidSpeech.cpp in Luigi/moss-transcribe-diarize-zhtw-gguf.

MOSS-Transcribe-Diarize zh-TW — quantized ONNX graphs

Quantized exports of Luigi/moss-transcribe-diarize-zhtw for on-device inference. Accuracy vs the bf16 fine-tune on held-out real meeting audio (script-normalized MER): int8 0.027–0.042, q4 0.068.

web/ — onnxruntime-web (browser) layout

3-graph pipeline with dynamic KV (past→present chaining), audio-token splice done by the caller: encoder.int8.onnx (mel[1,80,3000]→audio_embeds), embedding.int8.onnx (ids→embeds), decoder.{int8,q4}.onnx (inputs_embeds+attention_mask+past_k/v_i → logits+present_k/v_i), ecapa.onnx (wav[1,T]→L2-normed 192-d speaker embedding, STFT in-graph). Used by the fully-local demo: Luigi/zh-tw-transcriber-local (see its pipeline.js for reference inference code, incl. the exact Whisper-mel frontend and speaker-linking recipe).

sherpa/ — sherpa-onnx C++ layout

Fixed-KV-cache decoder interface (qwen3-asr style: input_ids + audio_features

• cache_position + fixed caches → logits + per-layer KV deltas) for the sherpa-onnx MOSS-SATS port. Tokenizer assets: use the tokenizer/ from the base model repo (vocab.json + merges.txt; inject added_tokens_decoder into tokenizer_config.json from tokenizer.json's added_tokens).

Example audio in validation: 立法院 IVOD open data (CC-BY-4.0).