Moonshine Voice

Moonshine Voice is an automatic speech recognition (ASR) solution developed by Useful Sensors, optimized for edge devices and real-time streaming applications, released in October 2024.

Key Advantages#

• Flexible Input Windows: No fixed 30-second window like Whisper, no zero-padding overhead
• Streaming Cache: Supports incremental audio input with encoder/decoder state caching
• Ultra-low Latency: Medium Streaming model achieves 107ms on MacBook Pro, ~802ms on Raspberry Pi 5
• Cross-platform Consistency: Unified API across iOS, Android, Linux, Windows, macOS

Performance Comparison (vs Whisper)#

Language Support#

English, Spanish, Chinese, Japanese, Korean, Vietnamese, Ukrainian, Arabic

Model Sizes#

Core Capabilities#

• Real-time Speech-to-Text (ASR)
• Voice Intent Recognition (semantic matching for predefined commands)
• Speaker Identification (distinguish different speakers)
• VAD Segmentation (based on Silero VAD)

Technical Architecture#

Microphone Input → VAD (Silero) → Streaming Encoder → Decoder → Text Output
                                    ↓
                               Speaker ID
                                    ↓
                               Intent Recognition

• Model Architecture: Encoder-Decoder Transformer
• Positional Encoding: Rotary Position Embedding (RoPE)
• Inference Engine: ONNX Runtime (.ort format, memory-mapped optimization)
• Quantization: 8-bit weights + 8-bit MatMul

Quick Start#

pip install moonshine-voice
python -m moonshine_voice.download --language en
python -m moonshine_voice.mic_transcriber --language en

Python API Example#

from moonshine_voice import Transcriber, TranscriptEventListener

transcriber = Transcriber(model_path=model_path, model_arch=model_arch)

class TestListener(TranscriptEventListener):
    def on_line_completed(self, event):
        print(f"Line completed: {event.line.text}")

transcriber.add_listener(TestListener())
transcriber.start()
transcriber.add_audio(audio_chunk, sample_rate)
transcriber.stop()

Key Configuration Options#

Use Cases#

• Real-time transcription apps (subtitles, meeting notes)
• Voice command systems (robotics, smart home, automotive)
• Edge device voice interaction (Raspberry Pi, IoT, wearables)
• Privacy-sensitive offline speech processing