Tether has released an open-source artificial intelligence model capable of transcribing speech directly on smartphones, smart glasses and humanoid robots without sending audio to cloud servers.
QVAC TranscribePsy contains 254 million parameters and can process 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max using Apple’s Core ML framework, according to an October 14 announcement from Tether AI Research. The company said it was faster than several competing speech-recognition systems in its tests.
The model supports 25 European languages on laptops and eight on phones. Its local operation means developers can use it without an internet connection, while versions optimised for Apple’s Neural Engine are available for compatible devices.
TranscribePsy is part of Tether’s QVAC AI project, which is designed to run AI applications on consumer hardware and keep user data private. On smart glasses, the software can process audio received through Bluetooth or Wi-Fi with minimal delay. Humanoid robots can use their onboard processors for transcription while retaining computing capacity for navigation, visual recognition and movement.
Tether said local speech processing could be particularly useful during medical consultations, legal interviews, therapy sessions and journalism, where recordings may contain confidential information. The system is designed to handle continuous audio and produce text shortly after speech is detected, supporting voice-controlled applications without transmitting recordings to external servers.
TranscribePsy performance
Tether reported a word error rate of 8.02% on the Open ASR Leaderboard, 6.58% on Multilingual LibriSpeech and 21.56% on Common Voice. Lower scores indicate fewer differences between the generated transcript and the original spoken words.
The company said TranscribePsy outperformed other systems in the same parameter category and compared it with larger models including Qwen3-ASR-0.6B, omniASR-CTC-1B-v2, OWSM-CTC-v4 and Whisper-Medium.
Despite having fewer parameters, it produced better results on Multilingual LibriSpeech than several models three to four times larger, according to Tether’s research team. Across six smartphones, it was 2.6 times faster than the next quickest system and 5.7 times faster than Whisper-small when running on phone CPUs.
The iPhone-optimised version used less than half the disk space and 1.7 times less memory than a competing Core ML model tested by the company. Tether published its evaluation configurations so researchers could reproduce the findings.
Four versions have been released for different hardware and language needs. A laptop version supports 25 European languages, while a smaller configuration covers eight commonly used languages. Mobile users can run the GGUF version through the parakeet.cpp engine in qvac-fabric-speech, while the Core ML edition is designed for Apple’s Neural Engine.
Paolo Ardoino, Tether’s CEO, said users should retain control over their conversations as voice-enabled devices become more widespread. He described TranscribePsy as an example of speech recognition that can run entirely on personal devices without centralised infrastructure.
Part of wider QVAC development
The release follows Tether’s QVAC SDK, introduced in April 2026 as an open-source toolkit for running AI applications on smartphones and computers. It supports iOS, Android, Windows, macOS and Linux, with tools for text generation, speech processing, image recognition and translation.
The software uses QVAC Fabric, a customised version of llama.cpp, alongside Holepunch technology for peer-to-peer model distribution and delegated processing. In March, Tether integrated BitNet LoRA into QVAC Fabric. The company said it had fine-tuned models of up to 3.8 billion parameters on the Pixel 9, Galaxy S25 and iPhone 16, and tested models of up to 13 billion parameters on the iPhone 16. GPU processing was between two and 11 times faster than CPU baselines in its evaluations.
In September, Ardoino disclosed progress on a peer-to-peer search engine operating across distributed nodes. Tether stored the complete Wikipedia archive across 100 nodes and demonstrated search without a centralised service. The system remains under development and is intended eventually to support thousands of interconnected nodes.
Earlier work included the QVAC Genesis I dataset, released in October 2025 with 41 billion synthetic tokens for training models in mathematics, science and technical reasoning, and QVAC Workbench, an application for running supported AI models locally.
TranscribePsy is available under the Apache 2.0 licence, with model weights, mobile GGUF builds, Apple Silicon Core ML versions and evaluation configurations. Within the QVAC Psy ecosystem, it is intended to work alongside VisionPsy, TranslatePsy-Nano and MedPsy. Tether identified document analysis, translation, navigation and private note-taking as possible future applications, while stating that those combined services remain prospective.
