Tether Releases 254 Million Parameter Offline Speech Model

Tether has released a 254 million parameter speech recognition model that can transcribe conversations directly on smartphones, smart glasses and humanoid robots without sending audio to cloud servers. According to an October 14 announcement from Tether AI Research, the open source model is called QVAC TranscribePsy.

The model transcribed 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max using Apple's Core ML framework. Tether stated that TranscribePsy performed faster than several competing speech recognition systems in its tests, operating 2.6 times faster than the next fastest system and 5.7 times faster than Whisper-small on phone CPUs.

Hardware and Language Support

TranscribePsy supports 25 European languages on laptops and eight on phones, with dedicated versions optimized for Apple’s Neural Engine. The model is released in four versions designed for different hardware and language requirements, including a GGUF version operating through the parakeet.cpp engine in qvac-fabric-speech.

For smart glasses, the software can process audio received over Bluetooth or Wi-Fi with minimal delay. Humanoid robots can run the processing locally on onboard processors. Tether noted that local processing protects confidential information in medical consultations, legal interviews, therapy sessions and journalism.

Performance Benchmarks

In performance evaluations released by Tether, TranscribePsy recorded an 8.02% word error rate on the Open ASR Leaderboard and 6.58% on Multilingual LibriSpeech. Its error rate on Common Voice was 21.56%. The model is available under the Apache 2.0 license, providing developers access to model weights, mobile GGUF builds, Apple Silicon Core ML versions and evaluation configurations.

Key facts

• Tether released a 254M parameter offline speech recognition model named TranscribePsy.
• The model transcribed 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max.
• TranscribePsy supports 25 European languages on laptops and eight on mobile devices.
• Performance benchmarks include an 8.02% word error rate on the Open ASR Leaderboard and 6.58% on Multilingual LibriSpeech.