Tether Releases Local AI Speech Model for Private Device Processing

The model, developed under Tether’s QVAC AI project, demonstrates high efficiency by transcribing 10 minutes of audio in just 1.65 seconds on an iPhone 17 Pro Max. Tether’s research indicates that the system outperforms significantly larger competitors, including Whisper-Medium, by maintaining a lower word error rate despite its compact size. Optimized versions for Apple’s Neural Engine and a GGUF format for mobile CPUs ensure compatibility across diverse hardware.

Beyond basic transcription, TranscribePsy supports up to 25 European languages and is engineered to run without an active internet connection. CEO Paolo Ardoino emphasized that the release aims to curb the reliance on centralized platforms for voice-enabled applications. By keeping data on-device, Tether intends to provide a secure alternative for developers building tools that handle confidential information. The project is available under an Apache 2.0 license, with future plans to integrate the model into a broader ecosystem of local AI applications, such as document analysis and image recognition.