Tether AI Research, the artificial intelligence group within Tether Data, released QVAC TranscribePsy on October 14, 2026. It is a 254-million-parameter automatic speech recognition (ASR) model designed to turn speech into text on laptops and phones without uploading audio to a server. Weights are available on Hugging Face under Apache 2.0, alongside GGUF builds for mobile CPUs, a Core ML build for Apple's Neural Engine and the evaluation configurations used for the published figures. Tether has not released its training code or its training data.
The family ships as four checkpoints. TranscribePsy-250M-EU handles 25 European languages and is aimed at laptops. TranscribePsy-250M-EU-sub narrows coverage to the eight languages used by Multilingual LibriSpeech (English, German, Dutch, French, Spanish, Italian, Portuguese and Polish) in exchange for better accuracy, and the GGUF and Core ML phone builds are derived from it. In the 25-language model, English runs on the main weights while a small adapter serves the remaining 24 languages; a language gate picks the path per utterance, which Tether says adds roughly 0.1 points of word error on English.
Under the hood it is a FastConformer-TDT transducer. Tether started from NVIDIA's 627-million-parameter Parakeet-TDT-0.6B-v3, shrank it through structured pruning and then trained the smaller network to imitate the original, using about 564,000 hours of speech in 25 languages. The team argues that beginning from a pruned copy of the teacher, rather than from random weights, is what lets a model this small stay competitive. A fuller technical report with per-language results has been promised but not yet published.
On Tether's own benchmarks, the 25-language model posts word error rates of 8.02% on the Open ASR Leaderboard suite, 21.56% on Common Voice and 6.58% on Multilingual LibriSpeech. For Whisper-small it reports 12.34%, 38.17% and 11.51% respectively. Against larger systems the picture is mixed: TranscribePsy leads Whisper-Medium, omniASR-CTC-1B-v2, OWSM-CTC-v4 and Qwen3-ASR-0.6B on Multilingual LibriSpeech, but Qwen3-ASR-0.6B is ahead on the Open ASR Leaderboard suite (7.40% against 8.02%) and omniASR-CTC-1B-v2 does better on Common Voice, with 20.07% versus TranscribePsy's 21.56%.
Speed is where Tether makes its strongest pitch. Across six handsets, from the iPhone 15 to the Pixel 10 and Galaxy S26 Ultra, each model transcribed the same 40 minutes of audio. The 8-language GGUF build averaged 1,021 milliseconds per utterance on the phone CPU, against 2,631 ms for omniASR-CTC-300M-v2 and 5,854 ms for Whisper-small, with a cold start near 252 ms and peak memory around 534 MiB. On an iPhone 17 Pro Max's Neural Engine, the Core ML build needed 1.65 seconds for a 10-minute recording after the model was loaded, although its cold start is 12.7 seconds according to the project's GitHub page. Android is slower: working through the full 40-minute set took the Galaxy S26 Ultra 62 seconds, while the Pixel 9 needed 180.
Tether frames the model as privacy infrastructure for devices that listen constantly, such as smart glasses and humanoid robots, and as the speech layer for its wider QVAC Psy line, which includes VisionPsy, TranslatePsy and MedPsy. For developers, the practical uses are offline dictation, meeting and lecture transcripts, captions and voice commands where connectivity is poor or audio should not leave the device.
Several limits apply. All accuracy and latency numbers come from Tether, and no independent replication was found at publication. Language coverage is European only, the current builds transcribe after the speaker finishes rather than streaming text live, and the release notes describe the Apache 2.0 weights as intended for research and educational purposes, a framing commercial adopters will want to clarify.