Moonshine Voice beats Whisper Large v3 with 7.5x fewer parameters

Moonshine Voice is a newly released open-weight speech-to-text (STT) model designed specifically for real-time on-device transcription. Developed by the six-person startup Moonshine AI, it runs locally across Python, iOS, Android, and Raspberry Pi without relying on cloud APIs.

What it can do

The standout achievement is its size-to-accuracy ratio. Its streaming model operates on just 200 million parameters—making it 7.5 times smaller than Whisper Large v3—yet it actually beats Whisper Large v3 on word error rate.

Beyond basic transcription, it supports English, Mandarin, Spanish, Japanese, Korean, and several other languages. The model also comes with built-in voice activity detection (VAD) and intent recognition for voice commands right out of the box.

Why it matters

If you build mobile apps, edge devices, or offline smart assistants on hardware like a Raspberry Pi, this is worth testing immediately. Running Whisper locally has always meant wrestling with high memory footprints and latency.

Moonshine shows that an optimized 200M model can match or exceed cloud-sized baselines on edge silicon while packing practical features like VAD and intent parsing. We think it marks a major step forward for truly usable, local-first voice interfaces.