- Tether AI Research unveiled QVAC TranscribePsy, a multilingual speech-recognition model that processes audio on devices without sending it to the cloud.
- According to Tether, the model’s Apple Core ML version transcribed 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max.
- The 254 million-parameter model supports 25 European languages on laptops and eight languages on mobile devices.
Tether AI Research, Tether’s research division, has unveiled QVAC TranscribePsy, a compact multilingual automatic speech-recognition model designed to run directly on devices. The release is expected to expand on-device artificial intelligence capabilities for smartphones, smart glasses and humanoid robots while protecting private conversations from being uploaded to external servers.
TranscribePsy has 254 million parameters and is designed to process audio in near real time. According to Tether, its Apple Core ML version can transcribe 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max.
Local processing and performance
Tether is positioning TranscribePsy as a tool for developers building voice AI services without cloud infrastructure. Local processing can help protect audio recorded at home, at work or in public, including medical consultations, legal interviews, psychotherapy sessions and journalists’ conversations with confidential sources.
Tether reported a word error rate, or WER, of 8.02% on the Open ASR Leaderboard, 6.58% on Multilingual LibriSpeech and 21.56% on Common Voice. A lower WER indicates fewer word-recognition errors.
On Multilingual LibriSpeech, TranscribePsy outperformed several models with three to four times as many parameters, according to the testing results. In tests across six smartphones, the model ran 2.6 times faster than its closest competitor and 5.7 times faster than Whisper-small based on phone CPU benchmarks.
The laptop version supports 25 European languages, while the mobile version supports eight. Tether released four specialized versions for different devices, including GGUF builds for mobile hardware and a Core ML build optimized for the iPhone’s Neural Engine.
Tether CEO Paolo Ardoino said:
“The future of AI shouldn’t mean giving centralized platforms a front-row seat to our personal lives. TranscribePsy shows that advanced speech AI can be fast, capable, and entirely local.“
Expansion of the QVAC ecosystem
TranscribePsy expands the QVAC ecosystem, through which Tether is developing tools to process text, images, video and other types of data directly on devices. The model may eventually integrate with VisionPsy, TranslatePsy-Nano and MedPsy for voice-based document analysis, translation, navigation and private notes.
In December 2025, Tether unveiled QVAC Genesis II, a synthetic educational dataset for AI training. After the addition of 107 billion tokens, the dataset reached 148 billion tokens across 19 educational domains.
In June 2026, the company updated QVAC SDK 0.13.0 with support for local video generation, brain-computer interfaces and robotics models.
TranscribePsy was released under the Apache 2.0 license for research and educational purposes. The release includes open model weights, mobile builds and configurations for reproducible testing.
The AI work forms part of a broader transformation at Tether. The company recently brought in KPMG to audit its reserves, after which Ardoino said Tether had long since stopped viewing itself solely as a crypto company.
Source: Incrypted
