Practical overview
When it fits
Use SenseVoice when an audio workflow needs transcription plus language, emotion, or event analysis, and the released checkpoint’s language coverage and input workflow fit the task.
Who it is for
- Developers building speech transcription or audio-analysis workflows.
- Teams that need language identification, emotion recognition, or audio-event tags alongside speech recognition.
- Researchers and practitioners adapting or fine-tuning speech models for domain-specific data.
Documented uses
- Transcribe speech in a supported language.
- Identify which supported language is spoken in an audio clip.
- Analyze speech for emotion and detect audio events.
- Explore a documented inference path or adapt the model with fine-tuning scripts.
Requirements & limits
The repository identifies the model as open source and lists an MIT license. The README specifies FunASR 1.3.26 or later for its examples.
- The five-language scope applies to the released SenseVoiceSmall checkpoint; the broader research scope reported by the project is not equivalent to checkpoint support.
- The documented direct-inference example accepts clips up to 30 seconds; longer recordings must be segmented, with VAD shown for that workflow.
- Speaker diarization is a separate composed FunASR pipeline using FSMN-VAD and CAM++, not an output of the SenseVoiceSmall checkpoint.
- The project notes that event classification has limitations compared with specialized audio-event-detection models.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run inference on a short audio clip
Use the documented direct-inference path to process an audio clip of up to 30 seconds.
Before you start
- FunASR 1.3.26 or later, as specified for the examples
- The SenseVoiceSmall checkpoint and an audio clip no longer than 30 seconds
- A CUDA device, as shown in the example
Steps
Load SenseVoiceSmall with its documented from_pretrained method and set the model to evaluation mode.
Pass the clip to the model's inference method, selecting a supported language or automatic language detection.
Apply the documented rich-transcription postprocessor to the returned text.
Expected outcome
A processed transcription result from the short audio clip; the model can also return its supported emotion and audio-event tags.
Guide limitations
- This direct-inference example is limited to input duration of 30 seconds or less.
- The README's example uses a CUDA device; it does not establish other hardware configurations.
Compare options for Model inference and serving
These projects document a common task and share a project role. This is a source-based comparison, not an execution benchmark or a guarantee of interchangeability.
| Project | Documented purpose | Platform | Access and requirements |
|---|---|---|---|
| Sensevoice | Transcribe speech in a supported language. | The repository identifies the model as open source and lists an MIT license. The README specifies FunASR 1.3.26 or later for its examples. | |
| Edgedict | Decode speech from a YouTube Live video into text. | Cloud |
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- QwenAudio/SenseVoice · Official homepage title · retained in its source language