pipecat-moonshine 0.1.0

Moonshine ASR (speech-to-text) community integration for Pipecat

pipecat-moonshine

Moonshine ASR speech-to-text integration for Pipecat.

Moonshine is a family of small, fast automatic-speech-recognition models optimized for resource-constrained devices. The Tiny English model is roughly 26 M parameters, the Base English model roughly 58 M — both run on CPU via ONNX Runtime with no GPU required. That makes Moonshine an attractive choice for low-latency, on-device transcription in Pipecat pipelines that already handle VAD upstream.

This package provides MoonshineSTTService, a SegmentedSTTService subclass that plugs straight into any Pipecat pipeline.

Status

Community-maintained integration. See Pipecat's Community Integrations guide for what that means — in short, the Pipecat team does not maintain or support this package; please file issues here.

Tested with Pipecat v1.2.1.

Installation

pip install pipecat-moonshine

This pulls in useful-moonshine-onnx and pipecat-ai automatically. The first time you instantiate the service, the chosen model weights are downloaded from Hugging Face (UsefulSensors/moonshine) and cached locally.

For the included foundational example you also need the local-audio extras:

pip install 'pipecat-moonshine[examples]'

Usage

from pipecat.pipeline.pipeline import Pipeline
from pipecat_moonshine import MoonshineSTTService, Model

stt = MoonshineSTTService(model=Model.TINY_EN)

pipeline = Pipeline([
    transport.input(),
    vad_processor,     # MUST run upstream of the STT — see below
    stt,
    # ... downstream processors
])

MoonshineSTTService subclasses SegmentedSTTService, so a VAD-driven processor (e.g. VADProcessor with SileroVADAnalyzer) must produce VADUserStartedSpeakingFrame / VADUserStoppedSpeakingFrame upstream of it. Each detected speech segment is decoded into a single final TranscriptionFrame — Moonshine does not emit interim results.

Configuring the model at runtime

Pass an explicit model or a fully-built settings object:

# Convenience kwarg
stt = MoonshineSTTService(model=Model.BASE_EN)

# Or via Settings (e.g. when you want to update at runtime)
stt = MoonshineSTTService(
    settings=MoonshineSTTService.Settings(model="moonshine/base"),
)

Running the example

git clone https://github.com/ubopod/pipecat-moonshine
cd pipecat-moonshine
pip install -e '.[examples]'
python examples/transcription-moonshine.py

Speak into your default mic; lines like Transcription: hello world will be printed for each detected utterance.

Audio format requirements

Moonshine expects 16 kHz, mono, 16-bit signed PCM input. Pipecat's default LocalAudioTransport and most WebRTC transports already provide this. If your pipeline runs at a different sample rate the service will log a warning on the first segment and transcription quality may degrade — add a resampler upstream if you need a different rate.

Moonshine also enforces a per-segment duration window: speech segments shorter than 0.1 s or longer than 64 s are silently dropped (the service logs at DEBUG level when this happens).

Supported models

Constant	Model name	Params	Notes
Model.TINY_EN	moonshine/tiny	26 M	English-only, MIT-licensed weights.
Model.BASE_EN	moonshine/base	58 M	English-only, MIT-licensed weights.

Multilingual models — important license note

Moonshine also publishes multilingual checkpoints (Spanish, Japanese, Arabic, Korean, Mandarin, Vietnamese, Ukrainian, …). Those weights are released under the Moonshine Community License, which is non-commercial.

For that reason they are intentionally not enumerated in the Model enum. If you want to use one you must:

1. Read and accept the upstream Moonshine Community License.

2. Pass the model name as a string explicitly, e.g.:

   stt = MoonshineSTTService(model="moonshine/base")
   # then load the appropriate language checkpoint via your own download flow
   

This package does not bundle, mirror, or auto-download non-commercial weights, and the maintainers make no representation that doing so complies with your use case.

Frames

In	Out
VADUserStartedSpeakingFrame	(no output — buffers audio internally)
VADUserStoppedSpeakingFrame	one TranscriptionFrame per segment (final), or nothing
Any non-VAD audio	buffered/forwarded according to SegmentedSTTService

Errors during transcription are pushed as ErrorFrames; the pipeline is not torn down so other services can continue.

Metrics

can_generate_metrics() returns True. Per-segment processing time is recorded via start_processing_metrics / stop_processing_metrics, so enabling metrics on your PipelineTask will surface Moonshine latency alongside the rest of your pipeline.

Maintainer

Community-maintained. Not affiliated with Moonshine AI or Daily.

License

BSD-2-Clause — see LICENSE. Note that the Moonshine model weights are governed by their own license (MIT for English models, Moonshine Community License for others) — see the section above.

Versioning and changelog

See CHANGELOG.md. This package follows semantic versioning.

Getting help

• Pipecat Discord: https://discord.gg/pipecat (#community-integrations)
• Pipecat changelog (track upstream changes that may affect this integration): https://github.com/pipecat-ai/pipecat/blob/main/CHANGELOG.md
• Issues for this integration: file them in this repo.