Transcribe Audio (Moonshine)

Summary: Ultra-lightweight speech-to-text optimized for speed and low resource usage, with hallucination mitigation via max-length capping.

Use This When

• You need the smallest possible ASR model footprint (27M–61M params)
• Running on constrained GPU memory alongside larger models
• Speed matters more than peak accuracy

What It Does

• Converts audio frames to text using UsefulSensors Moonshine (encoder-decoder architecture)
• Caps max token generation based on audio duration (6.5 tokens/sec) to limit hallucination on silence
• Supports variable-length input with no fixed 30-second window
• 565x realtime throughput on GPU

Works Best With

• Pipelines where GPU memory is tight and a tiny ASR model is needed
• Pair with detect-voice-activity to pre-filter silence before transcription
• Voice command interfaces where utterances are short

Caveats

• English-only (base and tiny variants)
• Autoregressive decoder can still hallucinate on silence despite max-length cap
• Best results when paired with upstream VAD filtering
• Model auto-downloads from HuggingFace on first run (~400MB for base)