Moonshine Streaming Medium

Moonshine is Useful Sensors' family of on-device ASR models optimized for live voice applications. The Streaming Medium checkpoint is the largest/most accurate streaming variant, part of the Moonshine Voice / Moonshine v2 release.

• Lightweight 50 Hz audio frontend (time-domain ops, CMVN, 2× causal stride-2 convs) — no mel-spectrogram
• Ergodic encoder: sliding-window self-attention with no positional embeddings, windows of (16,4) at edge layers and (16,0) in the middle for ~80 ms lookahead
• Adapter injects learned positional embeddings
• Causal Transformer decoder with RoPE, autoregressive
• Trained on ~300K hours of English speech (200K open + 100K internal)

• Only ~245M params, yet achieves 6.66% WER on OpenASR Leaderboard, beating Whisper Large v3 (7.44%, 1.55B params)
• True streaming: computes while the user is still speaking, enabling order-of-magnitude lower response latency than Whisper
• Runs on CPU-only, sub-1GB RAM, 0.1–1 TOPS — targets Raspberry Pi, phones, laptops
• Cross-platform C++ core (ONNX Runtime) with bindings for Python, Swift, Java, C++
• English-only by design (per 'Flavors of Moonshine' paper, monolingual models are more accurate at equal compute)

On-device voice agents, live captioning, accessibility tools, voice command/interface systems, edge/embedded deployments where Whisper is too large or too slow.