Transcribe Audio (Moonshine)

@moonshine1 VersionTranscribe

UsefulSensors Moonshine ASR. Resamples each AudioFrame to 16 kHz and decodes it to a String. Each checkpoint is monolingual, so the language is chosen by the model you load; for in-stream multilingual audio prefer transcribe_audio_faster_whisper.

How it fits

Typical backends

Tight-GPU voice command with VAD.

Streaming transcription into HTTP output.

Caveats

LimitationEach checkpoint understands one language; audio in another language is decoded as nonsense in the checkpoint's language without warning, so the `model` must match the spoken language.
Hard constraintThe `model` must name a real Moonshine checkpoint; an unknown id fails to load at startup.
AccuracyA decoder-length heuristic scales the output budget to the audio length to keep silence hallucinations short, but does not eliminate them; pair with detect_voice_activity_silero_vad upstream when long silences are common.
I/O contractThe single output is always a String, and empty input returns the empty string without running the model.
CostThe checkpoints are small (tens to a couple hundred million parameters), making this one of the lighter ASR options, though a GPU still helps latency.
State lifetimeConfiguration is read once at startup; changing it requires a restart.

Versions

7371e622latestdefault
Automated release
6/24/2026