Moonshine Open-Weights STT Outperforms Whisper Large V3
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Moonshine released second-generation open-weights STT models claiming higher accuracy than Whisper Large V3 at the top end, with the smallest models at 26MB for constrained deployments.
- The new models replace Whisper's fixed 30-second input window with flexible-length audio processing and add incremental caching to skip redundant computation during streaming.
- Moonshine trained language-specific models for Arabic, Japanese, Korean, Spanish, Ukrainian, Vietnamese, and Chinese, finding that single-language models achieve higher accuracy at the same size and compute.
- The first-generation Moonshine models already ran up to 5x faster than Whisper in live speech applications, according to the project.
- A cross-platform C++ core library using OnnxRuntime ships with native bindings for Python, Swift, Java, and C++, supporting iOS, Android, macOS, Linux, Windows, Raspberry Pi, and IoT devices.
- Whisper achieves sub-20% Word Error Rate on only 33 of its 82 supported languages at 1.5B parameters, and just 5 on the Base edge model, according to OpenAI data Moonshine cites.
Why it matters: For developers building real-time voice agents, Moonshine's flexible input windows and streaming cache directly target the sub-200ms latency bar that Whisper's 30-second fixed window makes hard to clear. The per-language training approach could unlock markets like Korea and Japan where the source notes Whisper's WER exceeds usable thresholds — only 5 of 82 languages hit sub-20% WER on Whisper's Base edge model. On-device execution across everything from Raspberry Pi to iOS also removes account and API-key dependencies for privacy-sensitive voice apps.
