Tether AI Research Releases On-Device Speech Recognition Model TranscribePsy, Targeting Smart Glasses and Humanoid Robots

October 14: Tether AI Research has released an open-source multilingual automatic speech recognition model, QVAC TranscribePsy, with approximately 254 million parameters, capable of running entirely on-device without the need to upload audio to the cloud. The company says it can process audio from smart glasses connected via Bluetooth or Wi-Fi in near real-time, and can also run on onboard processors of humanoid robots; on the Core ML version for iPhone 17 Pro Max, transcribing 10 minutes of audio takes about 1.65 seconds.

According to disclosed evaluation data, TranscribePsy has a word error rate (WER) of 8.02% on the Open ASR Leaderboard, 6.58% on Multilingual LibriSpeech, and 21.56% on Common Voice; it is the fastest among comparable models across six tested phones, about 2.6 times faster than the next-fastest model. The model covers 25 European languages on laptop and 8 European languages on mobile, and its model weights and builds such as GGUF and Core ML are released under the Apache 2.0 license. Tether CEO Paolo Ardoino said he hopes advanced voice AI will run locally on users' own devices, with data not having to leave the device.