Tether has released a 254 million parameter AI model that can transcribe conversations directly on smartphones, smart glasses and humanoid robots without sending audio to cloud servers.
The open-source model, called QVAC TranscribePsy, was announced on October 14 by Tether AI Research. It is designed to run on devices with limited computing power, allowing users to convert speech into text without an internet connection.
Tether said the model can transcribe 10 minutes of audio in just 1.65 seconds on an iPhone 17 Pro Max using Apple’s Core ML framework.
The model is part of Tether’s QVAC AI project, which focuses on running artificial intelligence directly on consumer devices. By processing audio locally, the system can reduce reliance on external servers and help keep sensitive conversations private.
The technology could be useful for smart glasses that receive audio through Bluetooth or Wi-Fi. Humanoid robots could also run it on their onboard processors while keeping computing resources available for tasks such as navigation, visual recognition and movement.
Tether said local speech recognition may be particularly useful in medical consultations, legal interviews, therapy sessions and journalism, where recordings can contain confidential information.
TranscribePsy delivers strong results in performance tests
Tether reported an 8.02% word error rate on the Open ASR Leaderboard and 6.58% on Multilingual LibriSpeech. Its score on Common Voice was 21.56%.
The company said TranscribePsy outperformed other speech recognition models in the same parameter category. It was also compared with larger systems, including Qwen3-ASR-0.6B, omniASR-CTC-1B-v2, OWSM-CTC-v4 and Whisper-Medium.
According to Tether’s research team, TranscribePsy performed better on the Multilingual LibriSpeech benchmark than several models three to four times its size.
Tests across six smartphones showed that the model ran 2.6 times faster than the next-fastest system and 5.7 times faster than Whisper-small when using phone CPUs.
The iPhone-optimized version also required less than half the disk space and 1.7 times less memory than a competing Core ML model included in the tests.
Word error rate measures how closely a transcription matches the original speech. A lower score means fewer mistakes. Tether also published its testing configurations so other researchers can reproduce the reported results.
The model supports multiple European languages
TranscribePsy comes in four versions designed for different devices and language requirements. One laptop version supports 25 European languages, while a smaller configuration covers eight languages on phones.
For mobile devices, Tether provides a GGUF version that runs through the parakeet.cpp engine in qvac-fabric-speech. A separate version uses Apple’s Core ML framework and is optimized for the Neural Engine in compatible devices.
The model is designed to process continuous audio and produce text shortly after speech is detected. This could help developers build voice-controlled applications that respond without sending recordings to outside servers.
Tether CEO Paolo Ardoino said the goal is to give users more control over their conversations as voice-enabled devices become increasingly common.
He said AI should not require people to give centralized platforms access to their personal lives. TranscribePsy is intended to show how speech recognition can work directly on personal devices without relying on centralized infrastructure.
Tether continues expanding its QVAC AI project
TranscribePsy follows several other releases under Tether’s QVAC project in 2026.
In April, Tether introduced the QVAC SDK, an open-source toolkit that allows developers to run 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 customized 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 to support training and running large AI models on consumer hardware. The company said it had fine-tuned models containing up to 3.8 billion parameters on devices such as the Pixel 9, Galaxy S25 and iPhone 16.
Tether also reported testing models with up to 13 billion parameters on the iPhone 16. In its evaluations, GPU processing delivered speed improvements of two to 11 times compared with CPU-based tests.
The company’s decentralized computing work extends beyond AI models. In September, Ardoino revealed progress on a peer-to-peer search engine designed to operate across distributed nodes. The team had stored the complete Wikipedia archive across 100 nodes and demonstrated search without relying on a centralized service. The project remains under development, with plans to support thousands of connected nodes.
Earlier initiatives include the QVAC Genesis I dataset, released in October 2025 with 41 billion synthetic tokens for training AI models in mathematics, science and technical reasoning. Tether also introduced QVAC Workbench, an application that allows users to run supported AI models locally.
Open-source release gives developers access to model weights
TranscribePsy is available under the Apache 2.0 license. Developers can access its model weights, mobile GGUF builds, Apple Silicon Core ML versions and evaluation configurations.
Within the QVAC ecosystem, the speech model is intended to work alongside other specialized systems, including VisionPsy, TranslatePsy-Nano and MedPsy.
Tether identified document analysis, translation, navigation and private note-taking as potential applications for future integrations. These are possible uses rather than a confirmation that all such features are already available through TranscribePsy.
The research team at Tether Data, S.A. de C.V., said its wider work focuses on developing compact AI models capable of processing audio, text, images and video directly on consumer devices.
