Desert Ant Labs released the Voz model for local speech recognition

Desert Ant Labs has introduced Voz — a specialized speech recognition model designed to work exclusively on Apple devices, utilizing the resources of the built-in Apple Neural Engine to ensure data processing privacy without relying on cloud servers.

The model supports transcription in 25 languages and provides time stamps at the individual word level, with a reported average word error rate (WER) of 7.40%, which decreases to 2.83% when working with longer audio recordings, thus increasing accuracy.

The architecture is optimized for the Apple ecosystem, allowing for seamless integration of voice recognition functionality into native applications without significant strain on the central processor, and the model is available for download on the Hugging Face platform, where developers can access weights and documentation for implementation in their own projects.

The release of Voz expands local audio processing capabilities for macOS and iOS developers, leveraging Neural Engine hardware acceleration to reduce power consumption during real-time transcription tasks, which is critical for mobile devices and laptops.