FunASR
Alibaba's industrial-grade speech toolkit: Paraformer, SenseVoice, and SeACo-Paraformer with hotword customization.
About
FunASR is the fundamental speech recognition toolkit from Alibaba's DAMO Academy Speech Lab, built to close the gap between ASR research and industrial deployment. Its flagship Paraformer is a non-autoregressive end-to-end model trained on 60,000 hours of manually annotated Mandarin, extended with timestamp prediction and hotword customization through SeACo-Paraformer. The toolkit also bundles an FSMN-based voice activity detector and a CT-Transformer punctuation restorer.
SenseVoice, the multilingual sibling, adds emotion recognition and audio event detection; the small variant transcribes Mandarin, Cantonese, English, Japanese, and Korean at a real-time factor of 0.007. The code is MIT, but the pretrained weights carry their own licenses, so check terms before commercial use. Alibaba also serves the models as paid realtime APIs on Bailian/DashScope for teams that would rather not self-host.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paraformer trained on 60,000 hours of manually annotated Mandarin (provider-reported, INTERSPEECH 2023 paper)
- SenseVoice-Small: RTF 0.007, 70 ms latency on 10 s audio (234M params, 5 languages) vs Whisper Large v3 RTF 0.111 (provider-reported, FunAudioLLM report)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Paraformer's 60,000-hour Mandarin training set and SenseVoice-Small's 0.007 RTF figure from the INTERSPEECH 2023 paper and the FunAudioLLM report; MIT code license confirmed in the repo; DashScope API availability documented by integrators. Strengths: industrial-scale training data behind open models, the fastest non-autoregressive ASR in its class, hotword customization built in. Docked for: (1) pretrained weights are subject to a separate model license with non-commercial terms on some weights, an anomaly worth flagging; (2) Mandarin-first design, with other languages carried mainly by SenseVoice. License: MIT (code); model weights under separate MODEL_LICENSE (verify commercial terms).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Paraformer trained on 60k hrs annotated Mandarin; SenseVoice-Small RTF 0.007 (70ms/10s audio)
- · Also served as paid realtime APIs on Alibaba Bailian/DashScope (fun-asr-realtime, sensevoice-realtime-v1)