FunASR

Quick Verdict

FunASR is an actively maintained end-to-end speech-recognition toolkit under the ModelScope organization. Its canonical repository is modelscope/FunASR, with official documentation at modelscope.github.io/FunASR. It is not one model. The project provides model loading, training and fine-tuning infrastructure, offline transcription, streaming WebSocket services, VAD, punctuation, speaker pipelines, an OpenAI-compatible server, MCP examples, and edge runtimes.

That distinction matters. Fun-ASR-Nano targets Chinese, English, Japanese, Chinese dialect groups, and regional accents; the separate MLT-Nano checkpoint covers 31 languages. SenseVoiceSmall combines five-language ASR with emotion and audio-event labels, while Paraformer supplies practical Chinese-English offline and low-latency streaming choices. Speaker diarization in the official pipeline comes from a separate CAM++ model, not from SenseVoice itself. As of July 20, 2026, the repository remained active and PyPI release notes had reached version 1.3.22.

FunASR is a strong fit for Chinese transcription and private deployment when a team can operate models and protect recordings. It is not automatically compliant merely because it is self-hosted, and the MIT code license does not govern every checkpoint. Organizations must review the model card and license for the exact weights they deploy.

Best For

• Engineering teams building private Chinese meeting, classroom, media, or contact-center transcription.
• Products requiring streaming recognition, incremental text, interruption handling, VAD, and controllable latency.
• Organizations that need CPU, GPU, edge, or isolated-network deployment instead of sending sensitive recordings to a cloud ASR vendor.
• Researchers who want swappable ASR, punctuation, VAD, speaker, emotion, and audio-event components in one toolkit.

For a ready-made meeting application rather than ASR infrastructure, compare Otter AI. If the task is speech generation rather than recognition, GPT-SoVITS and CosyVoice solve a different problem.

Key Features

• Chinese and multilingual ASR: Paraformer covers Chinese-English offline and streaming workloads. Nano and MLT-Nano have distinct language scopes and should be documented by exact model ID.
• Offline and streaming operation: process files through Python or CLI, or stream audio chunks through Paraformer and Nano-oriented real-time services.
• Composable long-audio pipeline: AutoModel can combine FSMN-VAD, CT-Transformer punctuation, and CAM++ speaker components to produce segments, timestamps, and speaker labels.
• SenseVoice metadata: SenseVoiceSmall recognizes speech in five languages and can emit emotion and audio-event labels such as laughter or applause. These fallible labels should support search or review, not high-stakes decisions.
• Hot-word support: contextual Paraformer models can boost product and domain vocabulary. Excessive boosting can also create substitutions, so evaluate recall and false positives together.
• Serving and agent integration: funasr-server exposes an OpenAI-compatible transcription endpoint, and official examples cover MCP and common application frameworks.
• CPU and edge runtime: llama.cpp/GGUF support and release binaries reduce Python dependencies for selected models. Quantization, architecture support, and accuracy must be tested on the target hardware.
• Current engineering surface: official materials include deployment matrices, security guidance, long-session diagnostics, subtitle output, Kubernetes examples, and client recipes.

Use Cases

For long-form Chinese audio, a production pipeline can use VAD to segment speech, run Paraformer or SenseVoice, add punctuation and timestamps, and use CAM++ to assign local speaker IDs. Human reviewers should be able to jump from text to the original time range and correct names, numbers, and domain terms. A diarization label such as “Speaker 2” does not establish real identity.

Live captions and voice assistants can use the streaming endpoints, but teams should benchmark end-to-end latency rather than model inference alone. Audio transport, VAD decisions, queueing, model decoding, punctuation, network jitter, and UI rendering all contribute. Contact-center automation should preserve confidence and traceability and must not let a probable transcript directly trigger a refund, account suspension, medical conclusion, or legal commitment.

Recordings and voice characteristics can be sensitive personal data. Before capture, confirm the right to process the recording or other media and disclose transcription, recording, speaker separation, retention, sharing, and any model-improvement use. Call-recording rules vary by jurisdiction and may require one-party or all-party consent; evaluate every participant’s applicable location rather than relying on a single global setting. FunASR recognizes speech but does not synthesize a voice or avatar, so an ordinary transcript should not be exaggerated as synthetic media. Where accuracy or speaker attribution could be misunderstood, identify the text as AI-transcribed and state whether it received human review. Never alter transcripts or misuse speaker IDs for unauthorized identity use, impersonation, fraud, or harassment. Voice agents should identify themselves and the recording practice at the start, offer an opt-out or human handoff, and use separate authentication for payment, account recovery, and identity-sensitive actions.

Pricing

The FunASR toolkit source is licensed under MIT and has no software subscription. Operating cost includes GPU or CPU capacity, storage, monitoring, annotation and correction labor, security, and incident response. Nano with vLLM is oriented toward GPU throughput; SenseVoiceSmall can serve CPU use cases; Paraformer streaming targets low-latency Chinese; and GGUF reduces runtime dependencies while introducing model and hardware tradeoffs.

Code and model licenses are separate. The official README states that pretrained weights use the license on each model card. Some cards point to the FunASR Model Open Source License Agreement 1.1, which includes attribution and model-name retention obligations; third-party checkpoints can use other terms. MIT on the loader does not make Qwen3-ASR, SenseVoice, Paraformer, community fine-tunes, or their training data MIT assets. Maintain a checkpoint inventory with source URL, hash, model card, license version, and approved use before deployment.

Pros

• Broad Chinese offline and streaming choices with GPU, CPU, and edge paths.
• VAD, punctuation, speaker, hot-word, and long-audio components are integrated under one toolkit.
• Python, CLI, WebSocket, OpenAI-compatible API, MCP, Docker, and GGUF entry points cover experimentation through deployment.
• ModelScope distribution is practical for users who prefer a China-accessible model source.
• The MIT codebase is inspectable and the canonical repository remained actively maintained as of the review date.
• Teams can choose components instead of accepting a single cloud model and feature bundle.

Cons

• Language coverage, latency, hardware needs, and auxiliary features are checkpoint-specific, making initial selection easy to misstate.
• High-throughput Nano/vLLM paths need suitable GPUs; quantized edge paths require a fresh accuracy and compatibility evaluation.
• Speaker, emotion, and event labels are probabilistic and unsafe as unreviewed inputs to employment, credit, medical, or identity decisions.
• MIT applies to repository code, while deployed weights can have custom and changing terms.
• Self-hosting does not supply consent records, tenant isolation, deletion workflows, monitoring, backups, or compliance evidence automatically.
• Accurate production transcripts still require representative evaluation data, domain vocabulary tuning, and human correction for critical content.

Alternatives

FAQ

Is FunASR a single model?

No. It is a toolkit and runtime for many ASR and supporting models. Selection, testing, deployment requirements, and licensing must be resolved for the exact checkpoint and pipeline components.

Does Fun-ASR-Nano support 31 languages?

The standard Nano checkpoint focuses on Chinese, English, Japanese, Chinese dialects, and regional accents. The separate Fun-ASR-MLT-Nano checkpoint is the documented 31-language choice. Record the full model ID in deployment documentation.

Does SenseVoice include speaker diarization?

SenseVoiceSmall itself provides ASR, emotion, and audio-event capabilities. The official combined example adds FSMN-VAD and CAM++ as separate models to produce speaker segments and IDs.

Does MIT cover all FunASR model weights?

No. MIT covers the toolkit repository. Each checkpoint is governed by its model card; some use the FunASR model agreement and third-party models may use other licenses.

Can speaker IDs be used for authentication?

No. Diarization labels separate conversational segments and do not verify a person’s legal or account identity. Use independent, preferably multi-factor authentication and route anomalies to trained staff.

What is required for compliant call transcription?

Disclose the AI, recording, purpose, retention, and sharing before capture; determine consent rules for the relevant jurisdictions; provide opt-out and human paths; minimize access and retention; and support correction and deletion where required.

Bottom Line

FunASR is valuable because it is a configurable speech-recognition toolkit, not because one benchmark number applies everywhere. Select a checkpoint for the actual language, latency, hardware, and auxiliary-feature requirements, then test it on representative recordings for accuracy, delay, concurrency, and error severity. Review every model license separately from the MIT code. With recording consent, AI disclosure, encryption, tenant isolation, deletion, human review, and safe escalation in place, FunASR can become a credible private ASR foundation rather than just a promising demo.