怎么返回时间戳

[gao12312]

Support returning timestamps 我看已经完成了，怎么usage

[haiewu]

参考demo2.py

[LauraGPT]

已支持。Fun-ASR-Nano 默认输出字级别时间戳：

from funasr import AutoModel

model = AutoModel(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    trust_remote_code=True,
    remote_code="./model.py",
    device="cuda:0",
    hub="hf",
)
res = model.generate(input=["audio.wav"], cache={}, batch_size=1, language="中文")
# timestamps 在结果中
print(res[0]["timestamps"])
# [{'token': '开', 'start_time': 0.42, 'end_time': 0.48}, ...]

配合 VAD 使用时时间戳会自动加上 VAD 偏移量。详见 README 的 Usage 部分。

[ititchuan]

"timestamps": [{"token": "そ", "start_time": 0.0, "end_time": 0.06, "score": 0.0}, {"token": "れ", "start_time": 0.06, "end_time": 0.12, "score": 0.0}, {"token": "で", "start_time": 0.12, "end_time": 0.18, "score": 0.0}, {"token": "は", "start_time": 0.18, "end_time": 0.24, "score": 0.0}, {"token": "お", "start_time": 0.24, "end_time": 0.3, "score": 0.0}, {"token": "願", "start_time": 0.3, "end_time": 0.36, "score": 0.0}, {"token": "い", "start_time": 0.36, "end_time": 0.42, "score": 0.0}, {"token": "し", "start_time": 0.42, "end_time": 0.48, "score": 0.0}, {"token": "ま", "start_time": 0.48, "end_time": 0.54, "score": 0.0}, {"token": "す", "start_time": 0.54, "end_time": 0.6, "score": 0.0}, {"token": "。", "start_time": 0.6, "end_time": 0.66, "score": 0.0}, {"token": "私", "start_time": 0.66, "end_time": 0.72, "score": 0.0}, {"token": "の", "start_time": 0.72, "end_time": 0.78, "score": 0.0}, {"token": "一", "start_time": 0.78, "end_time": 0.84, "score": 0.0}, {"token": "番", "start_time": 0.84, "end_time": 0.9, "score": 0.0}, {"token": "大", "start_time": 0.9, "end_time": 0.96, "score": 0.0}, {"token": "切", "start_time": 0.96, "end_time": 1.02, "score": 0.0}, {"token": "な", "start_time": 1.02, "end_time": 1.08, "score": 0.0}, {"token": "も", "start_time": 1.08, "end_time": 1.14, "score": 0.0}, {"token": "の", "start_time": 1.14, "end_time": 1.2, "score": 0.0}, {"token": "。", "start_time": 1.2, "end_time": 1.26, "score": 0.0}, {"token": "赤", "start_time": 1.26, "end_time": 1.32, "score": 0.0}, {"token": "い", "start_time": 1.32, "end_time": 1.38, "score": 0.0}, {"token": "ラ", "start_time": 1.38, "end_time": 1.44, "score": 0.0}, {"t...
[2026-07-21 09:57:58] INFO [modules.video_parser] [video_parser] [FunASR源文][CHUNK 01/06] [7.604-69.248] bytes=1972608 timestamps=254 hotword='' text='それではお願いします。私の一番大切なもの。赤いランドセル。私は赤が嫌いでした。子供の頃、両親が買ってくれるものはすべて。真っ赤な色をしていたからです。赤いパジャマ、赤い文房具、赤い髪飾り。と、周りのものはすべて赤。他の女の子が着ている水色のブラウスや。ピンクの髪飾りが羨ましくて、仕方ありませんでした。来日間もないあの日もそうでした。あの日、私は体験入学としてこれから通う日本の小学校に。胸を驚かせて言いました。今日、滝沢のおじいちゃんランドセルをくれるんだよね。と笑いながら聞く私に、母は微笑み返しました。'
时间戳返回的不对，不是音频真正的时间戳，而是根据 0.06s进行采样的结果，如何获取真的时间戳呢？

[LauraGPT]

重新打开。这个 Issue 在最初回答“如何返回时间戳”后关闭，但 @ititchuan 随后给出了新的、尚未解决的实测：返回值按固定 0.06s 递增，并非音频内容对应的真实对齐时间戳。这个后续不能因为原问题曾经关闭而被忽略。

当前先把状态准确标为“需要复现和定位”，不声称已修复。请补充以下最小信息，便于在相同输入上区分模型输出、VAD 偏移和后处理三个环节：

• funasr 版本、Fun-ASR-Nano 模型完整名称及 revision；
• 完整 AutoModel(...) / generate(...) 调用参数，特别是是否启用 VAD、batch/chunk；
• 一段可公开的短音频或最小复现音频，以及对应完整原始 JSON（不要只截取日志里的部分字段）；
• 期望的时间戳粒度：字级、词级还是句级。

我们会先在 exact revision 上复现固定步进的来源，再判断是模型能力边界、适配器错误还是 VAD/chunk 偏移没有正确回填。在有可复现结论或报告者确认前保持开放。

[LauraGPT]

I reproduced this with the exact same 5.616-second zh.mp3 sample and found that the fixed 60 ms spacing is not a frame-conversion bug. The Hugging Face checkpoint at revision 272c57b82523ada6fd87095e955f8e29100979ab contains none of the 86 configured CTC tensors, so the randomly initialized CTC branch was still being used for forced alignment.

I opened #164 with a fail-closed guard and four-language checkpoint documentation. Fresh end-to-end results on its exact head f341a2e412247f669283fcb764ae1cf4523492ed:

• HF old checkpoint (55ae0d2f...): transcription remains 开饭时间早上九点至下午五点。; CTC is disabled; no timestamps key is returned.
• Current ModelScope checkpoint (81fec861..., all 86 CTC tensors): the same text plus 14 timestamps from 0.78s to 5.16s, with 13 positive inter-token gaps and mean alignment score 0.90114.
• Repository tests: 28 passed; the commit is signed, DCO-compliant, and GitHub-verified.

The current workaround for native timestamps is hub="ms". The Hugging Face weights and remote code still need to be synchronized separately (tracked in modelscope/FunASR#3496).

I am keeping this issue open after the code merge. Please retest the published model/code path once it is available and confirm whether your original case is resolved; a merge alone is not treated as reporter verification.