vision-agents-plugins-moonshine 0.1.0

Moonshine STT integration for Vision Agents

Moonshine STT Plugin

This plugin provides Speech-to-Text functionality using Moonshine, a family of speech-to-text models optimized for fast and accurate automatic speech recognition (ASR) on resource-constrained devices.

Features

• Fast and Accurate: Moonshine processes 10-second audio segments 5x faster than Whisper while maintaining the same (or better!) WER
• Resource Efficient: Optimized for edge devices and resource-constrained environments
• Variable Length Processing: Compute requirements scale with input audio length (unlike Whisper's fixed 30-second chunks)
• Multiple Models: Support for both moonshine/tiny (~190MB) and moonshine/base (~400MB) models
• Device Flexibility: ONNX runtime automatically selects optimal execution provider
• Smart Sample Rate Handling: Automatic detection and high-quality resampling of WebRTC audio (48kHz → 16kHz)
• WebRTC Optimized: Seamless integration with Stream video calling infrastructure
• Efficient Model Loading: ONNX version loads models on-demand for optimal memory usage

Installation

From PyPI + GitHub (Required)

Since the Moonshine ONNX models are not available on PyPI, you need to install them separately from GitHub:

# 1. Install the core plugin from PyPI
pip install getstream-plugins-moonshine

# 2. Install the moonshine model dependency from GitHub
pip install "useful-moonshine-onnx @ git+https://github.com/usefulsensors/moonshine.git#subdirectory=moonshine-onnx"

With uv

# Install both dependencies
uv add getstream-plugins-moonshine
uv add "useful-moonshine-onnx @ git+https://github.com/usefulsensors/moonshine.git#subdirectory=moonshine-onnx"

Development Installation (uv)

If your project uses uv, add both dependencies to your pyproject.toml:

[project]
dependencies = [
    # … other deps …
    "getstream-plugins-moonshine>=0.1.0",
    "useful-moonshine-onnx @ git+https://github.com/usefulsensors/moonshine.git#subdirectory=moonshine-onnx",
]

[tool.uv.sources]
getstream-plugins-moonshine = { path = "getstream/plugins/moonshine" }  # for local development

Then:

uv sync        # installs both dependencies

Usage

from getstream.plugins.moonshine import MoonshineSTT
from getstream.video.rtc.track_util import PcmData

# Initialize with default settings (base model, 16kHz)
stt = MoonshineSTT()

# Or customize the configuration
stt = MoonshineSTT(
    model_name="moonshine/tiny",  # Use the smaller, faster model
    sample_rate=16000,            # Moonshine's native sample rate
    min_audio_length_ms=500,      # Minimum audio length for transcription
    # ONNX runtime will automatically select the best execution provider
)

# Set up event handlers
@stt.on("transcript")
async def on_transcript(text: str, user: any, metadata: dict):
    print(f"Final transcript: {text}")
    print(f"Confidence: {metadata.get('confidence', 'N/A')}")
    print(f"Processing time: {metadata.get('processing_time_ms', 'N/A')}ms")

@stt.on("error")
async def on_error(error: Exception):
    print(f"STT Error: {error}")

# Process audio data
pcm_data = PcmData(samples=audio_bytes, sample_rate=16000, format="s16")
await stt.process_audio(pcm_data)

# Clean up
await stt.close()

Model Selection

Moonshine offers two model variants with different trade-offs:

Model	Size	Parameters	Speed	Accuracy	Use Case
moonshine/tiny	~190MB	27M	Faster	Good	Resource-constrained devices, real-time applications
moonshine/base	~400MB	61M	Fast	Better	Default choice - balanced performance and accuracy

Default Model: The plugin uses moonshine/base by default as it provides the best balance of accuracy and performance for most use cases.

Choosing a Model:

• Use moonshine/tiny for maximum speed on very resource-constrained devices
• Use moonshine/base for better accuracy with still excellent performance (recommended)

Model Name Validation:

• Strict validation prevents silent fallbacks to wrong models
• Supports both short names ("tiny", "base") and full names ("moonshine/tiny", "moonshine/base")
• Clear error messages list all valid options when invalid model is specified
• Canonical model names ensure consistent behavior across different input formats

Sample Rate Handling

The Moonshine plugin automatically handles sample rate conversion for optimal transcription quality:

Events

The plugin emits the following events:

• transcript: Final transcription result

  • text (str): The transcribed text
  • user (any): User metadata passed to process_audio()
  • metadata (dict): Additional information including model name, duration, etc.

• error: Error during transcription

  • error (Exception): The error that occurred

Note: Unlike streaming STT services, Moonshine doesn't emit partial_transcript events as it processes complete audio chunks.