--- license: apache-2.0 base_model: openai/whisper-large-v3 language: - zh - nan pipeline_tag: automatic-speech-recognition library_name: transformers tags: - whisper - traditional-chinese - taiwan - taiwanese-hokkien --- # 羲和 HeliosLM ASR TW v6m 羲和(HeliosLM)ASR TW v6m 是以 **OpenAI Whisper large-v3** 為基礎、針對台灣華語(繁體中文輸出)與台語(漢字輸出)微調的語音辨識模型。它是羲和對話模型 [helioslm-chat-e4b](https://huggingface.co/chienhsinlin/helioslm-chat-e4b) 的語音搭配模型。 HeliosLM ASR TW v6m is a speech recognition model **fine-tuned from openai/whisper-large-v3** for Taiwanese Mandarin (Traditional Chinese output) and Taiwanese Hokkien (Han-character output). v6m 加強台語:台語漢字字錯率較 [v5](https://huggingface.co/chienhsinlin/helioslm-asr-tw-v5) 低,華語略高。以華語為主請用 v5,台語為主請用 v6m。 v6m improves Taiwanese Hokkien over v5; Mandarin is slightly worse. Use v5 for mostly-Mandarin audio, v6m for Taiwanese Hokkien. - 參數量 Parameters: about 1.55B (Whisper large-v3 architecture) - 授權 License: Apache-2.0 (same as the base model) ## 訓練資料 Training data All public data, no private recordings: - Mozilla Common Voice 27.0 zh-TW and nan-tw (CC0-1.0), test sentences and test speakers excluded - Google FLEURS cmn_hans_cn train/dev (CC-BY-4.0), transcripts converted to Traditional Chinese with OpenCC - AISHELL-1 train subset (Apache-2.0) Stage 1 (= v4): 1 epoch, learning rate 7e-6, 4,776 steps, then merged with the base model (0.85 × fine-tuned + 0.15 × whisper-large-v3) to give v5. Stage 2 (v6m): continued from v5 with Common Voice nan-tw (Han-character transcripts) upsampled 4×, 30% of the Common Voice zh-TW clips and FLEURS 2×; learning rate 4e-6, about 2.2 hours; then merged again (0.85 × stage 2 + 0.15 × whisper-large-v3). ## 評測 Evaluation (2026-11-05/06, self-run) Character error rate (CER, %, lower is better), greedy decoding, same clips and same run for both models. Scores are self-reported. | Test set | Clips | HeliosLM ASR TW v6m | HeliosLM ASR TW v5 | whisper-large-v3 (base) | | --- | --- | --- | --- | --- | | Common Voice 27.0 zh-TW test | 998 | 5.42 | 5.10 | 15.25 | | FLEURS cmn test (Traditional Chinese refs) | 945 | 7.61 | 7.33 | 7.54 | | Common Voice 27.0 nan-tw test, Han characters | 1,000 | 32.69 | 38.15 | 231.13 | | Common Voice 27.0 nan-tw test, raw transcripts (incl. romanization) | 1,000 | 85.09 | 86.29 | 121.57 | Taiwanese Hokkien is still weak: the model writes Han characters only and does not output Tâi-lô romanization. ## 使用 Usage ```python from transformers import pipeline asr = pipeline("automatic-speech-recognition", model="chienhsinlin/helioslm-asr-tw-v6m", chunk_length_s=30) print(asr("audio.wav", generate_kwargs={"language": "zh", "task": "transcribe"})["text"]) ``` 模型可能辨識錯誤,重要內容請自行確認。The model can make mistakes; verify important transcripts.