Transcribe Audio (SenseVoice)

@funaudiollm1 VersionTranscribe

FunAudioLLM SenseVoice transcriber. Resamples each AudioFrame to 16 kHz, decodes it, strips the event and emotion markers, and emits a clean String (empty on silence). For timestamped output prefer transcribe_audio_faster_whisper.

How it fits

Typical backends

Only speech segments reach the transcriber, whose multilingual text drives an LLM follow-up.

A recorded file is transcribed and the clean text posted as JSON.

Caveats

Hard constraintThe language setting accepts only its enumerated codes; any other value drifts to undefined internal behaviour and usually a degraded transcript.
Hard constraintThe model must name a real SenseVoice checkpoint; an unknown id fails to load at startup.
CompatibilityThe model loads custom code from its source repository at startup; pin the repository revision and cache it for production stability.
I/O contractThe post-processing strips the embedded event, emotion, and language markers (laughter, music, and similar) from the text, so they are not recoverable downstream; the single output is a String.
I/O contractEmpty input returns the empty string without running the model, and the model also returns the empty string when it detects silence.
CostFirst-run downloads are a few hundred megabytes; pre-seed the model cache for offline deployments. This is one of the lighter multilingual options.
State lifetimeConfiguration is read at startup and there is no runtime sync, so changing model, device, or language requires a restart.

Versions

7abdbc5elatestdefault
Automated release
7/8/2026