Fun-ASR-Nano + Transformers：先选对权重，再接入应用
+接入 Transformers，先选对权重
-在已有 Hugging Face 应用里接入语音识别，最容易踩的坑往往不是模型大小，而是把“同一模型家族”当成“同一种权重和接口”。Fun-ASR-Nano 已进入 Transformers 主线。这篇文章从 checkpoint、处理器到生成结果，解释如何把这次集成接到自己的应用中。
-先确认安装包，不只看合并状态
-PR #46180 于 2026-09-09 合并。当天核验的稳定版 5.16.1 还没有原生 fun_asr_nano 文件；源码包显示的 5.17.0.dev0 也不是下一稳定版发布承诺。请从固定版本接入指南安装明确的源码提交，不要只执行一次无版本约束的升级就假定已经兼容。
这种区别影响排障方向：未知模型类型可能是安装包未包含实现，不一定是 checkpoint 损坏；模型能下载也不代表所选运行时认识它。
如果你已经在用 Transformers，接入语音识别时先确认权重属于这条接口。同一个模型名字，不代表工具包、服务端和原生 Python 能互换文件。
+例如，文件下载完了，应用却提示“不认识模型类型”。先检查安装包有没有实现、权重是不是对应格式，不要急着把问题归为文件损坏。本篇选择官方 -hf 权重。
这条路径生成转写文本，不是现成服务。加载成功不等于拿到了字级时间戳、说话人分离或实时流式能力。
+ +
同一模型家族，四条不同路径
-| 应用已有的接口 | 选择什么 | 不要混淆 |
|---|---|---|
| FunASR AutoModel | Fun-ASR-Nano-2512 | 原始工具包路径；拆分引擎看对应文档。 |
| Transformers processor / generate | Fun-ASR-Nano-2512-hf | 本篇的原生接口，不是 vLLM 的转换包。 |
| 原生 vLLM HTTP 服务 | Fun-ASR-Nano-2512-vllm | 使用独立运行时、模型格式及服务参数。 |
| C++ / GGML 运行时 | 与运行时匹配的 GGUF | 不是把 Transformers 目录改后缀。 |
-
+
- FunASR AutoModel：
Fun-ASR-Nano-2512。原始工具包路径；拆分引擎看对应文档。
+ - Transformers processor / generate：
Fun-ASR-Nano-2512-hf。本篇的原生接口，不是 vLLM 的转换包。
+ - 原生 vLLM HTTP 服务：
Fun-ASR-Nano-2512-vllm。使用独立运行时、模型格式及服务参数。
+ - C++ / GGML 运行时：与运行时匹配的 GGUF。不是把 Transformers 目录改后缀。 +
这几个官方模型仓库位于 FunAudioLLM 下。原生 Transformers 路径加载 官方 -hf checkpoint，模型和处理器代码来自固定的 Transformers 安装，不需要执行 checkpoint 的远程 Python 代码。它减少了一种接入依赖，并不自动提供鉴权、队列或服务监控。
如果目标是多客户端并发请求，直接查看原生 vLLM 部署页；如果目标是离线 C++ 运行，查看llama.cpp 部署页。原生 Python 接口是一个可组合入口，不是所有服务的替换品。
从声音到文本，中间实际经过什么？
-- 音频样本。采样率与声道是数据含义的一部分。48 kHz 的数组不能只改参数标成 16 kHz；那会改变模型看到的时间尺度。指南先要求明确的单声道 16 kHz WAV。 -
- 音频特征。原生特征提取器用 torchaudio 的 Kaldi fbank，再做 LFR 堆叠与降采样。这里仍需要匹配版本的 torchaudio；不能套用 FunASR 工具包的可选依赖结论。 -
- 转写请求。
apply_transcription_request把音频、语言、上下文和关键词整理成模型的聊天模板，并对齐音频占位 token 与特征。
- - 生成和解码。
AutoModelForSpeechSeq2Seq返回 token。去掉输入 prompt 的长度后再解码，才能避免把模板混进转写文本。
先用合成静音验证配置和张量形状，可以把依赖、模板错误与权重加载问题分开。但预处理没有调用生成模型，不能据此宣布“识别通过”。
应用上下文应该放在哪里？
-业务词表通过原生 keywords 参数传入，相关背景通过 prompt 传入。它们不是工具包的 hotword 参数，也不是 HTTP 层字段。给两份录音分别传上下文时，语言、prompt 和嵌套关键词列表应与录音数量一致。输出顺序应回到同一份输入清单。
例如客服录音里的人名、产品名，可以作为候选关键词；但“模板中出现了关键词”和“模型在噪声中识别对了这个词”是两项测试。不要把拼写纠正或摘要改写后的文本冒充原始 ASR 输出。
-当前固定快照默认左侧 padding；批量示例仍可显式设置 padding 以便阅读。生成返回值要按整批输入张量的宽度去除 prompt，不是各条 attention mask 有效长度。空录音和空列表应在应用边界提前拒绝。
一次真实短录音检查说明了什么？
在固定源码、官方 -hf revision 和 CPU float32 环境中，我们运行了中文单录音、英文单录音、中英混合批量，以及一个带关键词的中文请求。四个请求均返回文本并在上限前生成 EOS；批量结果顺序正确。这是两份公开短音频的功能检查，不是准确率排行榜。
| 同一份中文音频 | 原始输出 |
|---|
| 参数 | 含义 | 不能保证什么 |
|---|---|---|
max_single_segment_time=30000 | FSMN-VAD 的端点阈值，单位毫秒（ms）。它参与帧级语音分段决策。 | 不是整文件时长限制，不保证严格等长片段或恰好切在一个词、句子的边界。 |
batch_size_s=300 | VAD 包装器的批预算，单位秒（s），内部转为毫秒；组批判据涉及最长片段时长乘以片段数量。 | 不是所有片段时长简单求和，也不是硬内存上限；较长的单段仍可能单独进入推理。 |
CPU 例外：当前包装器在 device="cpu" 时关闭上述组批预算，按片段逐个处理。因此调整示例中的 batch_size_s 不能当作 CPU 批吞吐优化已经生效。换用经过配置和验证的 GPU 环境时，应重新测量，而不是套用相同内存或速度结论。实现依据：批预算单位、CPU 分支与组批判据。
长文件上线前怎么验收
-
@@ -111,6 +97,26 @@
长文件上线前怎么验收
更多思路见会议转写验收清单与可复现性能评测。前者附带的 MOSS sentence_info 检查器有特定输入契约，不可直接套在任意 SenseVoice 返回值上；结构检查也不证明文字准确率或语音覆盖完整。
附录：窗口、资源与参数单位
+ +一次调用背后的资源路径
+这条离线路径会读取完整音频波形到 CPU 内存，再使用 VAD 区域组织 ASR 输入；它不是只从磁盘读取当前片段的有界内存流式处理。分段可以改变某一批推理的工作量，但整段波形、解码缓冲、分段信息、输出和并发请求仍占用资源。
+仅按 16 kHz、单声道、float32 波形计算，一小时数据约 230 MB；这只是数据量估算，不是进程 RAM 或 GPU 峰值内存实测，也不包含解码副本和模型。不能据此承诺任意文件都能处理，或一小时与一分钟文件显存相同。可接受的时长需要在所选硬件、模型、音频格式和并发条件下验证。
+ +两个参数，两个不同的边界
+-
+
max_single_segment_time=30000: FSMN-VAD 的端点阈值，单位毫秒（ms）。它参与帧级语音分段决策。 不是整文件时长限制，不保证严格等长片段或恰好切在一个词、句子的边界。
+batch_size_s=300: VAD 包装器的批预算，单位秒（s），内部转为毫秒；组批判据涉及最长片段时长乘以片段数量。 不是所有片段时长简单求和，也不是硬内存上限；较长的单段仍可能单独进入推理。
+
CPU 例外：当前包装器在 device="cpu" 时关闭上述组批预算，按片段逐个处理。因此调整示例中的 batch_size_s 不能当作 CPU 批吞吐优化已经生效。换用经过配置和验证的 GPU 环境时，应重新测量，而不是套用相同内存或速度结论。实现依据：批预算单位、CPU 分支与组批判据。
模型和服务应分开选择
- 本文是明确配置外部 FSMN-VAD 的 SenseVoice 离线示例。Paraformer 的配置与输出应另行核对；SenseVoice 富文本标签不是经过验证的情绪准确率。 @@ -118,32 +124,8 @@
- 需要实时麦克风转写时，按部署矩阵选择明确支持的流式模型与服务；缩小离线批次并不会把它变成流式协议。更多能力差异见模型选型指南。
模型和服务应分开选择
本文依据固定版本源码解释行为，没有新增长录音性能或准确率实测。旧页缺乏完整复现条件的耗时和全程覆盖宣传已撤下；历史版本保留在仓库中，不能作为当前部署承诺。
- -相关文章
--
-
- FunASR vs Whisper 实测对比 -
- SenseVoice 部署指南 -
- Fun-ASR-Nano 使用指南 -
- 说话人分离:谁在何时说话 -
- 情感与语种检测 -
- 实时流式语音识别 -
- 命令行转写(文本/JSON/SRT) -
- 自托管 OpenAI Whisper API 替代 -
- 自动生成字幕(SRT / VTT) -
- Python 语音转文字教程 -
- FunASR 跑进 llama.cpp(whisper.cpp 替代) -
- FunASR vs faster-whisper(中文/粤语) -
- 轻量语音识别(CPU 250MB) -
- 自托管替代 Deepgram/AssemblyAI -
- 选哪个 FunASR 模型 -
- 粤语语音识别(SenseVoice 原生粤语) -
- 日语语音识别(SenseVoice 转写+标点+情感) -
- Python 语音活动检测(VAD) -
- 自托管替代 Google/AWS/Azure 云语音 API -
- 中文语音识别(普通话)实战 -
- 标点恢复 Python 实战 -
- 语音识别带时间戳(字级) -
下一步，按会议转写验收清单挑一份有代表性的录音回听结尾。该文下载的 MOSS 检查器有不同输入契约，不能直接套用本篇 SenseVoice 返回值。
FunClip v2.2.0：用 MOSS 做长音频说话人分离与视频剪辑
- -
- FunClip v2.2.0 新增一条可选的 MOSS 路径：把长音频交给 vLLM 服务，FunASR 将模型输出归一化为文本、匿名说话人标签和时间段，FunClip 再生成 SRT 或按 spkS01、spkS02 剪辑。
MOSS-Transcribe-Diarize 是 OpenMOSS 维护的第三方模型，不属于 FunASR 或 FunClip。集成固定使用 OpenMOSS-Team/MOSS-Transcribe-Diarize revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8，并明确保留模型归属与支持边界。
spkS01 与 spkS02 只是当前录音内的匿名标签，不能识别已知人物、验证已注册声纹，也不保证不同录音中的同名标签对应同一个人。
数据路径
-| 阶段 | 职责 |
|---|---|
| vLLM | 加载固定 MOSS revision，通过 /v1/audio/transcriptions 返回带时间与说话人标记的 JSON。 |
| FunASR 1.4.9+ | 解析转写结果并生成统一的 text、timestamp 与 sentence_info。 |
| FunClip 2.2.0 | 渲染说话人 SRT、按说话人选段，或把时间段交给现有音视频剪辑流程。 |
1. 启动固定 revision 的 vLLM 服务
+把多人录音变成可剪辑的字幕
+ +如果你要从一段访谈里剪出某位嘉宾的发言，先得到能回到原声音的字幕，比先追求一整页流畅文字更有用。FunClip 可以用录音内的说话人标签选段；剪辑前仍要回听。
+已有的 FunClip v2.2.0 双人样例走通了这条链路：得到 S01/S02 两个标签、生成字幕，再按 S02 剪出片段。这是下文的历史功能案例，不是“任何访谈都能剪准”的证明。
+标签不是人物身份。spkS01 与 spkS02 只属于当前录音；换一段录音，同名标签不保证对应同一个人。段级字幕也不能代替逐字精确对齐。
先找到发言，再决定怎么剪
+MOSS-Transcribe-Diarize 是 OpenMOSS 维护的第三方模型。它提供转写和匿名说话人时间段；FunASR 整理结果，FunClip 把时间段接到字幕与音视频剪辑流程。
+- 给模型连续的录音。这条路径不接外部 VAD 或说话人模型，避免先切成独立请求后丢失录音内的说话人上下文。 +
- 在字幕里核对目标发言。标签帮助找到候选片段，但名字、数字、抢话和转场仍要回听，不能凭标签数量验收。 +
- 先导出一段，再批量处理。回看实际片段的开头、结尾和字幕；不要只检查剪辑命令退出成功。
哪些剪辑适合这条路径？
+- 按整段或说话人剪辑：MOSS 的段级时间戳可作为入口，FunClip 也支持有效的不足一秒短说话片段。 +
- 任意文字的精确字符级剪辑：保留 Paraformer 路径，并核对所选模型的语言和时间戳能力，不要把一句话平均切成字时间。 +
- 最后一段没有结束时间戳：发布记录中的 FunClip 会明确报错并提示提高
--moss-max-tokens，而非静默丢弃该段。调整上限后仍需核对尾部，不能据此保证完整召回。
开始新部署时，用维护中的 MOSS 部署指南核对运行时、显存和服务格式。下方旧命令仅解释这次发布，不替代当前环境准备；也不要把服务直接暴露到公网。
+附录：v2.2.0 历史发布记录
+以下保存原发布的命令、归档散列与测试口径，不是当前安装配方。其中未固定版本的 pip install -U vllm 不能保证今天的兼容性；不要直接复制升级。依赖、服务别名和请求格式应按维护指南重新核对，FunClip 命令的相对路径要求位于对应仓库根目录。
历史集成固定使用 OpenMOSS-Team/MOSS-Transcribe-Diarize revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8。当时的 FunASR 1.4.9+ 集成整理 text、timestamp 和 sentence_info，FunClip 2.2.0 渲染 SRT 并剪辑；不是所有未来版本的兼容承诺。
当时的服务与请求
python -m venv .venv-moss
. .venv-moss/bin/activate
pip install -U vllm
@@ -47,41 +53,29 @@ 1. 启动固定 revision 的 vLLM 服务
--revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8 \
--served-model-name moss-transcribe-diarize \
--trust-remote-code --host 127.0.0.1 --port 8898先用真实音频检查服务契约。当前经过验证的格式是 response_format=json：
这次历史验证使用 response_format=json，不是对所有 MOSS 后端格式的通用说明：
curl -fsS http://127.0.0.1:8898/v1/audio/transcriptions \
-F file=@sample.wav \
-F model=moss-transcribe-diarize \
-F response_format=json \
-F max_completion_tokens=81922. 启动 FunClip
+当时的 FunClip 启动参数
python -m pip install -U -r requirements.txt
python funclip/launch.py \
--model moss \
--moss-backend vllm \
--moss-base-url http://127.0.0.1:8898/v1 \
--moss-max-tokens 8192远端服务需要 bearer token 时，把凭据放进 MOSS_API_KEY 环境变量；FunClip 不要求把 token 写进命令行或仓库。
能力与边界
--
-
- 支持长音频 ASR、匿名说话人标注、SRT，以及按说话人剪辑，包括不足一秒的有效短说话片段。 -
- MOSS 提供段级时间戳，适合按整段或说话人剪辑；任意文本的精确字符级剪辑仍应使用 Paraformer。 -
- 单次录音内的标签一致性依赖模型看到连续音频，因此不接外部 VAD 或说话人模型，避免预切块破坏标签一致性。 -
- 如果最后一个 MOSS 片段因 token 上限没有结束时间戳，FunClip 会明确报错并提示提高
--moss-max-tokens，不会静默丢掉尾段。
-
完整生产部署、健康检查和容量边界见 MOSS 双语部署指南。
- -下载与校验 v2.2.0
+需要 bearer token 时通过 MOSS_API_KEY 环境变量提供，不把凭据放进命令行或仓库。
归档与功能检查
FunClip-2.2.0.tar.gz
SHA256 994c5d9cf392b74b36284d526eca8bada1560a3e7825ab7baa9c673a1b4ef216
FunClip-2.2.0.zip
SHA256 4f5a7d33d9ea65467f29b55b15ed5be18e64de7e57e2f9f36fe51a23b40557e7从 FunClip v2.2.0 Release 下载归档和 SHA256SUMS，不要从不明镜像复制二进制或凭据。
c205bf32a8b11226ff5e8acb9a3c7a1f00cd3b06。仓库测试结果为 84 passed、1 skipped；H100 + vLLM 两说话人样例返回 S01/S02、有效 SRT，并成功按 S02 剪辑。这个验证证明发布链路可用，不代表所有语言、音频条件或并发规模。先按部署指南跑通一段真实双人音频，再进入 FunClip 处理自己的视频。
查看 FunClip v2.2.0 发布与下载归档和 SHA256SUMS 来自 FunClip v2.2.0 Release。对应 commit c205bf32a8b11226ff5e8acb9a3c7a1f00cd3b06 的历史仓库测试为 84 passed、1 skipped；H100 + vLLM 双人样例返回 S01/S02、有效 SRT，并按 S02 剪辑。这不是本次编辑重跑的结果，也不代表所有语言、音频条件或并发规模。
下一步：从一段获授权的短访谈开始，按MOSS 部署指南跑通同一条链路，并人工回看第一个导出的片段。
HTTP 200 不等于转写成功：会议录音的验收清单
+会议转写，怎样才算做好了？
-一次请求返回了文本，离“这份会议记录可以交给同事”还有多远？把 ASR 接进字幕、会议纪要或按说话人剪辑时，需要分别验收内容、说话人、时间和处理完整性。下面用 FunASR 的 MOSS 结果讲清这些层次，并提供一个只依赖 Python 标准库的结构检查脚本。
+团队准备交付会议纪要或字幕时，需要确认名字、决策和原声音能对得上，而不只是“接口返回了文本”。自动检查可以找出结构问题，交付质量仍需要回听和人工确认。
+下文的合成案例是一份12秒录音结果，最后一段到第8秒结束。剩下4秒值得回听，却不能直接算成“漏识别了4秒”：那里也可能只有静音。
+检查器没有听过录音。两个说话人标签、或接近终点的时间戳，都不等于转写正确；录音内的匿名标签也不是人物身份。
先分清四个问题
| 检查层次 | 要回答的问题 | 不能用什么代替 |
|---|
| Your application interface | Artifact | Important distinction |
|---|---|---|
| FunASR AutoModel | Fun-ASR-Nano-2512 | Original toolkit path; split-engine has its own guide. |
| Transformers processor / generate | Fun-ASR-Nano-2512-hf | The native path described here, not a vLLM conversion. |
| Native vLLM HTTP service | Fun-ASR-Nano-2512-vllm | Separate runtime, checkpoint format and service options. |
| C++ / GGML runtime | Matching converted GGUF files | Not a renamed Transformers directory. |
-
+
- FunASR AutoModel:
Fun-ASR-Nano-2512. Original toolkit path; split-engine has its own guide.
+ - Transformers processor / generate:
Fun-ASR-Nano-2512-hf. The native path described here, not a vLLM conversion.
+ - Native vLLM HTTP service:
Fun-ASR-Nano-2512-vllm. Separate runtime, checkpoint format and service options.
+ - C++ / GGML runtime: Matching converted GGUF files. Not a renamed Transformers directory. +
The official model repositories are under FunAudioLLM. Native Transformers loads the official -hf checkpoint, while the model and processor code come from the pinned Transformers installation. It does not need to execute checkpoint-provided remote Python code. That removes one integration dependency; it does not supply authentication, queues or service monitoring.
For concurrent clients, inspect the native vLLM deployment. For offline C++, inspect llama.cpp deployment. A composable Python interface is not a replacement for every serving stack.
What happens between audio and text?
-- Audio samples. Sample rate and channel layout are part of the data's meaning. Labeling a 48 kHz array as 16 kHz changes the time scale seen by the model. The guide requires an explicitly prepared mono 16 kHz WAV. -
- Audio features. The native extractor computes Kaldi fbank through torchaudio, then applies low-frame-rate stacking and subsampling. Matching torchaudio is required here, regardless of the toolkit's optional-dependency policy. -
- The transcription request.
apply_transcription_requestprepares the checkpoint's chat template from audio, language, context and keywords, aligning audio placeholder tokens with features.
- - Generation and decoding.
AutoModelForSpeechSeq2Seqreturns tokens. Remove the input prompt width before decoding to avoid mixing template framing into recognized speech.
A synthetic-silence preprocessing check separates dependency and template failures from weight loading. It does not run the generation model, so it cannot establish successful speech recognition.
Where does application context belong?
-Pass a vocabulary through native keywords and relevant background through prompt. These are not the toolkit's hotword argument or HTTP fields. For separate recording contexts, language, prompt and nested keyword lists must match the number of audio inputs. Map results back to the same input manifest in order.
Customer-service recordings may benefit from candidate product or person names. But “the template includes this keyword” and “the recognizer recovered it in noise” are different tests. Do not present a spelling correction or summary rewrite as raw ASR output.
-The pinned checkpoint defaults to left padding; explicit padding still makes a batch example easier to inspect. Remove the full input tensor width from generated sequences, not each row's valid attention-mask length. Reject empty recordings and empty batches at your application boundary.
What did the short-recording check establish?
With pinned source, official -hf revision and CPU float32, we ran a Chinese recording, an English recording, a mixed Chinese/English batch and a Chinese keyword request. All four requests returned text and EOS before the limit, with the batch in input order. These are functional checks on two public short recordings, not an accuracy ranking.
| The same Chinese audio | Raw output |
|---|
| Parameter | Meaning | What it does not guarantee |
|---|---|---|
max_single_segment_time=30000 | An FSMN-VAD endpoint threshold in milliseconds (ms), used in frame-level speech segmentation decisions. | It is not a file-duration limit or a guarantee of exactly equal segments or cuts at word and sentence boundaries. |
batch_size_s=300 | A VAD-wrapper batch budget in seconds (s), converted to milliseconds internally. Grouping considers the longest segment duration multiplied by the number of segments. | It is not simply the sum of segment durations or a hard memory cap; a longer segment may still be processed alone. |
CPU exception: the current wrapper disables that grouping budget for device="cpu" and handles segments individually. Changing batch_size_s in this example is therefore not evidence of CPU batch-throughput optimization. A prepared and validated GPU setup needs new measurements, not inherited memory or speed conclusions. See budget units and the CPU branch and grouping rule.
Acceptance before admitting long recordings
-
@@ -111,6 +97,26 @@
Acceptance before admitting long recordings
See the meeting-transcript acceptance checklist and reproducible performance guide. The former's MOSS sentence_info audit script has a specific input contract, not an adapter for arbitrary SenseVoice results. Structural checks do not prove text accuracy or complete speech coverage.
Appendix: windows, resources and parameter units
+ +The resource path inside a file call
+This offline path loads the complete audio waveform into CPU memory, then organizes ASR inputs using VAD regions. It is not a bounded-memory stream that reads only the current segment from disk. Segmentation can change the work in one inference batch, but the full waveform, decoding buffers, segment metadata, outputs and concurrent requests still consume resources.
+For a 16 kHz, mono, float32 waveform alone, one hour is about 230 MB. This is a data-size estimate, not measured peak process RAM or GPU memory, and excludes decoder copies and model state. It cannot establish support for every file or equal GPU memory for one-hour and one-minute inputs. Validate acceptable durations on the chosen hardware, model, audio format and concurrency.
+See whole-file loading and segment sorting and the audio-decoding entry point. Source inspection explains the data path; it does not replace peak-memory measurement.
+Two parameters with different boundaries
+-
+
max_single_segment_time=30000: An FSMN-VAD endpoint threshold in milliseconds (ms), used in frame-level speech segmentation decisions. It is not a file-duration limit or a guarantee of exactly equal segments or cuts at word and sentence boundaries.
+batch_size_s=300: A VAD-wrapper batch budget in seconds (s), converted to milliseconds internally. Grouping considers the longest segment duration multiplied by the number of segments. It is not simply the sum of segment durations or a hard memory cap; a longer segment may still be processed alone.
+
CPU exception: the current wrapper disables that grouping budget for device="cpu" and handles segments individually. Changing batch_size_s in this example is therefore not evidence of CPU batch-throughput optimization. A prepared and validated GPU setup needs new measurements, not inherited memory or speed conclusions. See budget units and the CPU branch and grouping rule.
Choose models and services separately
- This is an offline SenseVoice example with explicitly configured external FSMN-VAD. Check Paraformer configuration and outputs separately. SenseVoice rich-text tags are not verified emotion accuracy. @@ -118,32 +124,8 @@
- For realtime microphone transcription, select an explicitly supported streaming model and service using the deployment matrix. Smaller offline batches do not create a streaming protocol. See model selection for capability differences.
Choose models and services separately
This article explains pinned source behavior, not new long-recording performance or accuracy measurements. The previous page's timing and full-coverage promotion lacked complete reproduction conditions and has been removed. Repository history preserves that text; it is not a current deployment promise.
- -Related posts
--
-
- FunASR vs Whisper Benchmark -
- SenseVoice Deployment Guide -
- Fun-ASR-Nano Guide -
- Speaker Diarization: Who Spoke When -
- Emotion & Language Detection -
- Real-Time Streaming Speech-to-Text -
- Transcribe from the Command Line -
- Self-Hosted OpenAI Whisper API Alternative -
- Auto-Generate Subtitles (SRT / VTT) -
- Speech to Text in Python -
- FunASR on llama.cpp (whisper.cpp Alternative) -
- FunASR vs faster-whisper (Chinese/Cantonese) -
- Lightweight Speech Recognition on CPU -
- Self-Hosted Deepgram/AssemblyAI Alternative -
- Which FunASR Model? -
- Cantonese Speech Recognition (SenseVoice) -
- Japanese Speech Recognition (SenseVoice) -
- Voice Activity Detection in Python -
- Self-Hosted Google/AWS/Azure STT Alternative -
- Chinese Speech Recognition in Python -
- Punctuation Restoration in Python -
- Speech-to-Text with Timestamps in Python -
Next, use the meeting-transcript checklist to choose what to listen for at the end of a representative recording. Its downloadable MOSS checker has a different input contract from this SenseVoice recipe.
FunClip v2.2.0: Long-Form Speaker-Aware Video Clipping with MOSS
- -
- FunClip v2.2.0 adds an opt-in MOSS path: send long audio to a vLLM service, normalize text, anonymous speaker labels, and time ranges through FunASR, then generate SRT or clip by spkS01, spkS02, and later speaker labels.
MOSS-Transcribe-Diarize is a third-party model maintained by OpenMOSS, not a FunASR or FunClip-owned checkpoint. The integration pins OpenMOSS-Team/MOSS-Transcribe-Diarize revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8 and keeps ownership and support boundaries explicit.
spkS01 and spkS02 distinguish anonymous speakers within the current recording. The model does not identify a known person, verify an enrolled voiceprint, or promise that labels match across separate recordings.
Data path
-| Stage | Responsibility |
|---|---|
| vLLM | Loads the pinned MOSS revision and serves timestamped speaker markup through /v1/audio/transcriptions. |
| FunASR 1.4.9+ | Parses the response into common text, timestamp, and sentence_info fields. |
| FunClip 2.2.0 | Renders speaker-labeled SRT and sends selected speaker ranges into the existing audio/video clipping workflow. |
1. Start vLLM at the pinned revision
+Turn a conversation into editable subtitles
+ +If you need one guest's remarks from an interview, subtitles that lead back to the sound are more useful than a page of fluent text. FunClip can select ranges by recording-local speaker labels; you still need to listen before accepting the cut.
+An existing FunClip v2.2.0 two-speaker example completed that sequence: obtain S01/S02 labels, produce subtitles, then cut the S02 ranges. This historical functional case anchors the article; it does not prove that every interview will be edited correctly.
+A recording-local label does not identify a known person. spkS01 and spkS02 belong to this recording, with no promise of the same person behind a label in another file. Segment subtitles are not precise word alignment.
Find the remark before choosing the cut
+MOSS-Transcribe-Diarize is a third-party model maintained by OpenMOSS. It supplies transcription and anonymous speaker time ranges; FunASR normalizes the result, and FunClip connects the ranges to subtitles and media editing.
+- Keep the recording context. This path uses no external VAD or speaker model. Splitting it into independent requests first can lose the context needed for consistent recording-local labels. +
- Check the target remarks in the subtitles. Labels locate candidate passages. Listen to names, numbers, interruptions and transitions rather than treating a label count as acceptance. +
- Export one cut before scaling up. Replay its start, end and subtitles. A successful editing command is not enough.
Which edits fit this path?
+- Whole-segment or speaker-based cuts: MOSS provides segment-level timestamps, and FunClip also supports valid speaker turns shorter than a second. +
- Precise arbitrary character-level text edits: keep the Paraformer path and check its language and timing support. Dividing a sentence evenly does not create word alignment. +
- A final segment without its end timestamp: the documented FunClip release reports truncation and suggests increasing
--moss-max-tokens, instead of silently discarding that segment. Check the tail again; a larger limit does not guarantee complete recall.
For a new deployment, use the maintained MOSS guide to check runtime, memory and response format. The old commands below explain this release, not current environment preparation. Do not expose the backend directly to the public network.
+Appendix: the historical v2.2.0 release
+The original commands, archive hashes and verification scope follow. They are not a current installation recipe. In particular, the unpinned pip install -U vllm is not a compatibility guarantee today; do not copy it as an upgrade instruction. Recheck dependencies, served model names and formats in the maintained guide. Relative FunClip commands assume the corresponding repository root.
The historical integration used OpenMOSS-Team/MOSS-Transcribe-Diarize revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8. Its FunASR 1.4.9+ integration normalized text, timestamp and sentence_info; FunClip 2.2.0 rendered SRT and performed the cuts. This is not a compatibility promise for every future version.
The service and request used then
python -m venv .venv-moss
. .venv-moss/bin/activate
pip install -U vllm
@@ -47,41 +53,29 @@ 1. Start vLLM at the pinned revision
--revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8 \
--served-model-name moss-transcribe-diarize \
--trust-remote-code --host 127.0.0.1 --port 8898Verify the service with real audio before starting FunClip. The tested contract uses response_format=json:
This historical check used response_format=json; it is not a universal format contract for every MOSS backend:
curl -fsS http://127.0.0.1:8898/v1/audio/transcriptions \
-F file=@sample.wav \
-F model=moss-transcribe-diarize \
-F response_format=json \
-F max_completion_tokens=81922. Start FunClip
+The FunClip launch options used then
python -m pip install -U -r requirements.txt
python funclip/launch.py \
--model moss \
--moss-backend vllm \
--moss-base-url http://127.0.0.1:8898/v1 \
--moss-max-tokens 8192For an authenticated remote service, put the bearer credential in the MOSS_API_KEY environment variable. FunClip does not require putting the token in the command line or repository.
Capabilities and boundaries
--
-
- Long-form ASR, anonymous speaker labeling, SRT, and speaker-based clipping are supported, including valid speaker turns shorter than one second. -
- MOSS provides segment-level timestamps. Use it for whole-segment or speaker clipping; keep Paraformer for precise arbitrary character-level text clipping. -
- Consistent speaker assignment within one recording depends on continuous context, so the MOSS path attaches no external VAD or speaker model that would pre-chunk the recording. -
- If the final segment is missing its ending timestamp after token exhaustion, FunClip raises an explicit truncation error and asks for a higher
--moss-max-tokensvalue instead of silently dropping the tail.
-
See the bilingual MOSS production guide for health checks, runtime choices, and capacity boundaries.
- -Download and verify v2.2.0
+When bearer authentication is required, provide it through MOSS_API_KEY, not command arguments or the repository.
Archives and functional checks
FunClip-2.2.0.tar.gz
SHA256 994c5d9cf392b74b36284d526eca8bada1560a3e7825ab7baa9c673a1b4ef216
FunClip-2.2.0.zip
SHA256 4f5a7d33d9ea65467f29b55b15ed5be18e64de7e57e2f9f36fe51a23b40557e7Download the archives and SHA256SUMS from the FunClip v2.2.0 release.
c205bf32a8b11226ff5e8acb9a3c7a1f00cd3b06. The repository suite completed with 84 passed and 1 skipped. A live H100 + vLLM two-speaker run returned S01/S02, valid SRT, and the expected S02 clip. This verifies the release path, not every language, audio condition, or production concurrency level.Run one real two-speaker recording through the production guide, then bring the same service into FunClip.
Open the FunClip v2.2.0 release and downloadsArchives and SHA256SUMS are in the FunClip v2.2.0 release. Commit c205bf32a8b11226ff5e8acb9a3c7a1f00cd3b06 had 84 passed and 1 skipped repository tests; the H100 + vLLM two-speaker case returned S01/S02, valid SRT and an S02 cut. These are historical results, not reruns for this edit or evidence for every language, recording condition or concurrency level.
Next, use one short interview you are authorized to process. Follow the MOSS deployment guide through the same path, then replay the first exported cut.
HTTP 200 is not a usable transcript: a meeting-audio acceptance checklist
+When is a meeting transcript ready to use?
-A transcription request returned text. Is it ready for meeting notes, subtitles or speaker-based editing? Treat transport, words, speakers and timing as separate acceptance questions. This article uses a FunASR MOSS result and a small standard-library Python checker to show what can be automated, and what still needs listening.
+Before your team shares meeting notes or subtitles, check whether readers can trust the names, decisions and links back to the sound. A successful request is only the start; structure checks and listening answer different questions.
+In the synthetic example below, a 12-second recording has segments ending at second 8. The four-second tail is a reason to inspect the audio, not proof of four seconds of missing speech: it may contain silence.
+A checker cannot hear the recording. Neither two speaker labels nor a timestamp at the end establishes a correct transcript, and recording-local labels do not identify known people.
Four questions, four kinds of evidence
| Layer | Question | Not a substitute |
|---|
