1. FunASR 是什么：它解决的不是“识别”，而是“整条链路”

1. What is FunASR: Solving the "Pipeline", Not Just "Recognition"

官方 README 对 FunASR 的定位很直接：它希望在语音识别的学术研究和工业应用之间搭桥。 所以你会发现它不是只给你一个 checkpoint，然后让你自己拼剩下的一切；相反，它把工程里常见的一整条链路都显式放到了工具层里。

The official README describes FunASR explicitly as a bridge between academic speech research and industrial application. Instead of giving you a raw checkpoint and leaving you to build the rest, it explicitly integrates the entire common engineering pipeline into the toolkit layer.

2. 核心原理：统一接口背后，其实是一条可组合的语音流水线

2. Core Principles: A Composable Speech Pipeline Behind a Unified API

官方示例里最常见的写法是 from funasr import AutoModel。这背后的设计哲学很清楚： 把“模型选择”和“链路编排”收敛到一个入口，让用户可以按需把 VAD、ASR、标点、说话人、时间戳拼起来。

The most common idiom in the official examples is from funasr import AutoModel. The design philosophy is clear: Converge "model selection" and "pipeline orchestration" into a single entry point, allowing users to combine VAD, ASR, Punctuation, Diarization, and Timestamps on demand.

2.1 长音频为什么先过 VAD

2.1 Why Long Audio Goes Through VAD First

启用 vad_model 后，VAD 会先把长音频切成更短的片段，再送给 ASR。这有两个直接收益：规避显存爆炸/高延迟，以及让离线模型能处理无尽的长播客。

When vad_model is enabled, VAD cuts long audio into shorter segments before feeding them to the ASR. This yields two benefits: avoiding OOM/high latency, and allowing offline models to process infinitely long podcasts.

2.2 流式识别的关键：Chunk 与 Cache

2.2 The Key to Streaming: Chunk and Cache

在 streaming 示例里，FunASR 使用 chunk_size 和 cache 等参数组织推理。它并不是简单切片，而是显式建模流式上下文：chunk_size 决定出字粒度和未来视野，而 cache 保留历史状态。

In streaming examples, FunASR uses parameters like chunk_size and cache. It doesn't just slice audio; it explicitly models streaming context: chunk_size dictates emission granularity and look-ahead, while cache preserves historical state.

3. 模型家族：分工明确的组件矩阵

3. Model Family: A Matrix of Specialized Components

FunASR 把不同任务拆成了多个可搭配的模型，你可以按场景裁剪能力和成本。

FunASR splits tasks into multiple composable models, letting you tailor capabilities and costs to your scenario.

4. 安装与上手

4. Installation & Getting Started

pip3 install -U funasr modelscope huggingface_hub

4.1 一行代码构建完整流水线

4.1 Complete Pipeline in One Line

from funasr import AutoModel
model = AutoModel(model="paraformer-zh", vad_model="fsmn-vad", punc_model="ct-punc")
res = model.generate(input="audio.wav", batch_size_s=300, hotword="魔搭")

4.2 SenseVoice 语音理解示例

4.2 SenseVoice Speech Understanding Example

model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad")
res = model.generate(input="audio.mp3", language="auto", use_itn=True)

5. 流式与部署 (Runtime & ONNX)

5. Streaming & Deployment (Runtime & ONNX)

FunASR 提供了专门的 Runtime 库用于生产部署，支持 C++ SDK、Docker、高并发调度，并支持导出为 ONNX 以实现跨平台 CPU 推理。

FunASR provides a dedicated Runtime repository for production deployment, supporting C++ SDKs, Docker, high-concurrency scheduling, and ONNX exports for cross-platform CPU inference.

# 导出 ONNX / Export to ONNX
funasr-export ++model=paraformer ++quantize=false ++device=cpu

5.1 流式 chunk 演示器

5.1 Streaming Chunk Demo

下面这个交互块不是在跑真实声学模型，而是在概念层面模拟 FunASR 流式推理时的窗口移动。 你可以把蓝色块理解为已经进入历史缓存的 chunk，把高亮框理解为当前推理窗口，把输出区理解为持续累积的 partial text。

The interactive block below does not run a real acoustic model; it simulates the concept of how a FunASR streaming window advances. Think of the blue blocks as chunks that have already entered history cache, the highlighted frame as the current inference window, and the output area as incrementally accumulated partial text.

5.2 我会怎么理解部署选型

5.2 How I think about deployment choices

• 如果你只是离线批量转写，Python + AutoModel 往往已经够用。
• 如果你需要高并发、稳定服务、镜像化交付，应该尽早转向 runtime SDK / docker。
• 如果你的目标平台偏 CPU、跨平台或推理引擎统一，ONNX 路线通常更自然。

• If you only need offline batch transcription, Python + AutoModel is often enough.
• If you need concurrency, service stability, and containerized delivery, move early toward runtime SDK / docker.
• If your target platform is CPU-heavy, cross-platform, or uses a unified inference engine, ONNX is often the more natural path.

6. 场景模型推荐器（交互式）

6. Scenario Model Recommender (Interactive)

根据您的具体需求，选择最合适的 FunASR 组件组合。点击下方选项卡查看详细架构与参数建议：

Select the most appropriate FunASR component combination based on your specific requirements. Click the tabs below to view detailed architectures and parameter recommendations:

model = AutoModel(model="paraformer-zh", vad_model="fsmn-vad", punc_model="ct-punc", spk_model="cam++")

model = AutoModel(model="paraformer-zh-streaming")
res = model.generate(input=chunk, cache=cache, is_final=is_final, chunk_size=[0, 10, 5])

model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad")
res = model.generate(input="audio.wav", language="auto", use_itn=True)

model = AutoModel(model="paraformer-zh", spk_model="cam++", vad_model="fsmn-vad")

7. 优势与局限

7. Advantages and Limitations

8. 总结

8. Conclusion

如果你的目标是处理长音频、生成带时间戳和标点的字幕、或者需要高并发的在线服务部署，FunASR 是目前最成熟的开源选择之一。它最大的价值在于将散落的论文能力，拧成了一股工业可用的绳。

If your goal is handling long audio, generating timestamped/punctuated subtitles, or deploying high-concurrency online services, FunASR is one of the most mature open-source choices. Its true value lies in weaving scattered academic capabilities into an industrial-strength rope.