Transcribe Audio (Parakeet)

@nvidia1 SürümTranscribe

NVIDIA Parakeet / Nemotron (NeMo) transcriber. Resamples each AudioFrame to 16 kHz and emits oneof t[String, [Subtitle], {text: String, language: String}] — plain text, timed segments, or text with the language it was recognised as. mode picks whole-frame transcription or cache-aware streaming, chunk_ms sets the streaming latency, and language conditions a multilingual checkpoint or asks it to detect the language itself.

How it fits

Typical backends

A recorded file is transcribed to timed subtitles and posted as JSON.

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

A live mixed-language meeting is captioned as it is spoken, each caption carrying the language it was recognised as. Set mode to 'streaming', chunk_ms to the latency the room can tolerate, and leave language on 'auto'. The language arrives with the frame that closes an utterance and labels the frames before it, so the step that assembles a caption holds the utterance's fragments until that frame.

Caveats

Hard constraintThe `model` must name a compatible NeMo ASR checkpoint; an unknown id fails to load at startup.
Hard constraintmode `streaming` needs a checkpoint whose encoder does cache-aware streaming. One that does not is refused at startup instead of being run frame by frame under a streaming name.
LimitationWhich languages are recognised is a property of the checkpoint, not of this component. A single-language checkpoint accepts only language `auto` and reports no language; a prompt-conditioned multilingual checkpoint takes a language-locale key or detects one.
I/O contractThe plain-text arm returns one String; the subtitle arm returns per-segment start/end times; the record arm returns the text with the language-locale it was recognised as, which is the empty string until the model has reported one. Empty input returns the empty result for the selected arm without running the model.
I/O contractUnder mode `streaming` a frame emits only the text finalised while it was consumed, so a frame that completes no chunk emits the empty result. Subtitle times then come from the stream clock, counting from the first frame, rather than from within one frame.
State lifetimeUnder mode `streaming` the encoder cache and the partial hypothesis persist for the life of the container, and the hypothesis grows with the stream. Restart the deployment between unrelated recordings so a new stream does not continue the previous one.
CostThe pretrained weights are a couple of gigabytes and the toolkit pulls a heavy dependency tree, so images are larger than the Whisper or Moonshine alternatives; pre-seed the model cache for offline deployments.
CompatibilityLicensing follows the bound checkpoint; confirm compatibility with your deployment terms before shipping one.
State lifetimeConfiguration is read once at startup; changing it requires a restart.

Sürümler

bd174c31latestdefault
Automated release
16.09.2026