Home
Latest reporting
News
Tether says its 254M-parameter TranscribePsy model runs locally on phones, glasses and robots, with benchmark claims and open-source weights.
Tether AI Research on 14 October 2026 said it has released QVAC TranscribePsy, an open-source speech recognition model designed to run entirely on-device rather than sending audio to cloud servers. The company said the model is aimed at smartphones, smart glasses and humanoid robots, and framed the launch as part of its broader push for private, voice-enabled AI.
Evidence and context
According to Tether’s release, TranscribePsy is a 254-million-parameter multilingual automatic speech recognition model that can process audio locally, including from smart glasses connected via Bluetooth or Wi-Fi and from a humanoid robot’s onboard processor. Tether said this approach is intended to reduce latency and avoid uploading spoken conversations that may contain sensitive personal, workplace or medical information.
The company also reported benchmark results for the model, saying TranscribePsy recorded an 8.02% word error rate on the Open ASR Leaderboard, 6.58% on Multilingual LibriSpeech and 21.56% on Common Voice. Tether said those results led all tested models in its parameter class, but those claims come from the company’s own announcement and should be read as reported benchmarks rather than independently verified comparisons.
Tether said the model’s Core ML build transcribes 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max, while using less than half the disk space and 1.7 times less memory than a tested Core ML competitor. It also said that across six tested smartphones, TranscribePsy was the fastest model, running 2.6 times faster than the next-fastest model and 5.7 times faster than Whisper-small on phone CPUs. The release did not provide the underlying test conditions in the material supplied here.
The announcement sits alongside a run of recent Tether AI releases. Verified background material shows the company unveiled Genesis III on 23 September 2026 to train AI to explain, not just answer, STEM problems, and released open-source translation models for African and European languages on 2 September 2026. Tether’s QVAC initiative, which it positions as a local-first AI ecosystem, also previously described a brain-to-text model running fully on-device in July 2026.
What the release includes
Tether said TranscribePsy ships with four specialized checkpoints: a laptop model covering 25 European languages, a laptop variant focused on eight core languages, a GGUF mobile build running on the qvac-fabric-speech parakeet.cpp engine, and a Core ML build optimized for Apple’s Neural Engine. The company said the model is released under the Apache 2.0 license for research and educational purposes, with open weights and evaluation configurations included.
The company also said the model is intended as the speech-to-text layer for possible future integrations within the QVAC Psy ecosystem, including VisionPsy, TranslatePsy-Nano and MedPsy. Tether described those possible uses as voice-driven document analysis, translation, navigation and private note-taking, but those downstream applications were presented as prospective integrations rather than confirmed product launches.
Paolo Ardoino, Tether’s chief executive, said in the release that the future of AI should not require centralized platforms to access personal conversations, and argued that speech AI can be run locally on devices users already own. The company’s broader QVAC materials, verified as of 14 October 2026, describe a local-first, peer-to-peer AI framework built around user-controlled hardware rather than centralized cloud systems.
The main limitation of the announcement is that its performance and efficiency claims are company-reported and were not independently corroborated in the supplied material. No commercial pricing, rollout schedule or next verified catalyst was provided.
Be the first to comment