Tether releases an open-weight speech model that transcribes audio offline

Tether AI Research released QVAC TranscribePsy on October 14, 2026, giving developers an open-weight speech-recognition model that runs on phones and laptops without uploading recordings. The 254-million-parameter model supports eight languages on phones and 25 on laptops. Its immediate promise is private, offline transcription—not live text appearing while someone is still speaking.

TranscribePsy began with NVIDIA’s Parakeet-TDT-0.6B-v3, a speech model with 627 million parameters. Tether reduced its size through structured pruning, removing parts of the larger model. It then used that original model as a teacher, training the smaller system to reproduce its outputs rather than learning from transcripts alone.

Tether says training covered about 564,000 hours of speech across 25 languages. Starting with a pruned version preserves knowledge already learned by the teacher, according to the company. Learning from the teacher’s outputs is also intended to pass that behavior to languages with less training data.

The downloadable weights are available under Apache 2.0, alongside evaluation configurations. That gives developers the model and a way to reproduce its reported tests, but not the whole training process: training code and training data are not included.

1.65 secondsTime to transcribe a 10-minute recording

Tether reports this result on an iPhone 17 Pro Max using the eight-language Core ML build, after the model was loaded.

The release offers two language configurations and two on-device builds. Laptops can run either language model; phones use the eight-language version. It covers Dutch, English, French, German, Italian, Polish, Portuguese and Spanish. The larger configuration adds 17 European languages, including Ukrainian, Greek and Finnish.

Two routes onto a phone

• GGUF build: runs the eight-language model on the phone’s CPU, its general-purpose processor, through the qvac-transcribepsy engine.
• Core ML build: uses the iPhone’s Neural Engine, the part of its chip designed for AI, through the same engine.

There is an important boundary for voice-app developers: this version transcribes a recording or utterance only after it finishes. Streaming, which would produce text during speech, remains in development. Tether also says it is working on smaller, faster models and more languages, with a fuller technical report and per-language results to follow.

Tether’s phone evaluation used six devices spanning two iPhones, two Samsung Galaxy phones and two Google Pixels. Each processed the same 160 utterances: 40 minutes of audio across eight languages. The published figures average the best three of five runs, a method Tether says keeps heat-related slowing between runs out of the numbers.

Across phone CPUs, the company reports an average response time of 1,021 milliseconds per utterance. That compares with 2,631 milliseconds for omniASR-CTC-300M-v2, its next-fastest comparison, and 5,854 milliseconds for Whisper-small. Peak memory use stayed between 530 and 542 MiB across the six devices in those tests.

Accuracy is measured by word error rate: substitutions, omissions and added words per 100 words in a reference transcript. Lower is better. Tether reports 6.58% on Multilingual LibriSpeech, averaged across eight languages, compared with 11.51% for Whisper-small and 7.76% for Whisper-Medium under its evaluation pipeline.

Larger competitors still won on other test sets. Qwen3-ASR-0.6B scored 7.40% against TranscribePsy’s 8.02% on the Open ASR Leaderboard. On Common Voice, omniASR-CTC-1B-v2 scored 20.07% against 21.56%. Those company-run comparisons support a competitive compact model, not an across-the-board accuracy lead.

For confidential work, the practical distinction is where transcription happens. Once downloaded, TranscribePsy can turn an interview, consultation or meeting recording into text on the device itself. Tether also identifies offline voice commands and captions for recorded talks as uses that do not require an internet connection.