rift-local

Local inference server for RIFT Transcription. Serves streaming speech recognition over WebSocket, backed by local models with automatic download.

Install

pip install rift-local

Backend extras

rift-local supports multiple ASR backends, each installed as an optional extra:

pip install rift-local[sherpa]      # sherpa-onnx (Nemotron, Kroko)
pip install rift-local[moonshine]   # Moonshine Gen 2 (via moonshine-voice)
pip install rift-local[sherpa,moonshine]  # both

On Apple Silicon, add MLX support for future GPU-accelerated batch transcription:

pip install rift-local[mlx]

For development (includes pytest):

pip install rift-local[dev]

Models

List all available models and see which are installed:

rift-local list
rift-local list --installed

sherpa-onnx models

Model	Params	Languages	Download	Notes
nemotron-en	0.6B	EN	447 MB	Best accuracy.
zipformer-en-kroko	~30M	EN	55 MB	Lightweight, fast. Only ~68 MB on disk.

Requires: pip install rift-local[sherpa]

Moonshine models

Model	Params	Languages	Size	Notes
moonshine-en-tiny	34M	EN	26 MB	Fastest. Good for low-resource.
moonshine-en-small	123M	EN	95 MB	Balanced speed/accuracy.
moonshine-en-medium	245M	EN	190 MB	Default. Best Moonshine accuracy.

Requires: pip install rift-local[moonshine]

Moonshine models are downloaded automatically by the moonshine-voice library on first use.

Usage

Server mode (for RIFT app)

Start the WebSocket server with any model:

# Start server and open RIFT Transcription in your browser
rift-local serve --open

# Moonshine (default model)
rift-local serve

# sherpa-onnx
rift-local serve --model nemotron-en

# Custom host/port
rift-local serve --model moonshine-en-tiny --host 0.0.0.0 --port 8080

The --open flag launches RIFT Transcription in your browser, pre-configured to connect to the local server. The voice source is set to "Local" automatically — just click to start the mic.

For local development of the RIFT Transcription client:

rift-local serve --open dev          # opens http://localhost:5173
rift-local serve --open dev:3000     # custom port

The server auto-downloads the model on first run, then listens on:

• WebSocket: ws://127.0.0.1:2177/ws (streaming ASR)
• HTTP: http://127.0.0.1:2177/info (model metadata)

Server options

Flag	Default	Description
--model	moonshine-en-medium	Model name from registry
--host	127.0.0.1	Bind address
--port	2177	Server port
--threads	2	Inference threads
--open	off	Open browser to RIFT Transcription client

WebSocket protocol

1. Client connects to /ws
2. Server sends info JSON (model name, features, sample rate)
3. Client sends binary frames of Float32 PCM audio at 16 kHz
4. Server sends result JSON messages with partial/final transcriptions
5. Client sends text "Done" to end the session

Running tests

# Install dev + backend dependencies
pip install -e ".[dev,sherpa,moonshine]"

# Run fast tests (mocked backends, no model download)
pytest

# Run all tests including slow integration tests (downloads models)
pytest --slow

Tests are in the tests/ directory:

• test_server.py — WebSocket server tests using a mock backend
• test_moonshine.py — Moonshine adapter unit tests (mocked) + integration tests (slow)
• conftest.py — Shared MockBackend fixture and --slow flag

Spec

See specs/rift-local.md for the full design document.

License

MIT