Introduction

On October 14, Tether AI Research released QVAC TranscribePsy, a 254M-parameter open automatic speech recognition model built to run on phones and laptops — not in a data center. Weights ship under Apache 2.0. The laptop cut covers 25 European languages; the phone sub-model covers 8, with GGUF and Core ML builds ready the same day.

This is the voice layer of QVAC’s on-device Psy stack: private dictation, meeting notes, offline captions, and voice controls that keep audio on the device. It is not streaming yet — you transcribe a finished utterance or recording, then get text.

Primary sources: the QVAC TranscribePsy blog (Models · October 13/14, updated October 14), the Hugging Face technical write-up (October 14), and the HF collection. Optional press context: Tether’s on-device speech note.

What Shipped

  • Architecture: 254M FastConformer-TDT, structured-pruned and distilled from NVIDIA’s Parakeet-TDT-0.6B-v3 (~627M teacher), trained on about 564,000 hours across the 25 languages.
  • License: open weights, Apache 2.0 (evaluation configs included; training code and data are not).
  • Language cuts:
    • TranscribePsy-250M-EU — 25 languages (Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Ukrainian). English on main weights; small adapters for the other 24; optional language gate (~0.1% WER cost on English when language is unknown).
    • TranscribePsy-250M-EU-sub — 8 Multilingual LibriSpeech languages (Dutch, English, French, German, Italian, Polish, Portuguese, Spanish) for higher accuracy when you only need that set.
  • On-device builds: GGUF (Q8_0) for phone CPUs and Core ML for the iPhone Neural Engine — both based on the 8-language sub model — plus full-precision checkpoints for laptops via NVIDIA NeMo.
  • Engine: tether-ai-research/qvac-transcribepsy; QVAC recommends qvac-fabric-speech.cpp’s parakeet engine for the fastest on-device path.
  • Limitation: batch / utterance transcription only — streaming is in progress, not day-0.

Accuracy and Phone Speed

Against models in the 242–370M class, TranscribePsy posts the lowest word error rate on all three suites QVAC reports. On Multilingual LibriSpeech (8-language mean), TranscribePsy-250M-EU lands at 6.58% WER — roughly half the errors of Whisper-small (11.51), OWSM v4 small (11.84), and omniASR-CTC-300M-v2 (11.89) under the same Open ASR Leaderboard normalizer. It also beats several models three to four times larger on MLS; on other suites a larger model sometimes wins (for example Qwen3-ASR-0.6B on the Open ASR Leaderboard mean).

On phones, QVAC timed the 8-language build on six devices (iPhone 17 Pro Max, iPhone 15, Galaxy S26/S25 Ultra, Pixel 10/9). The Core ML cut on an iPhone 17 Pro Max Neural Engine turns a single 10-minute recording into text in 1.65 seconds once loaded (~363× real time). Averaged across phone CPUs, the GGUF build answers an utterance in about 1,021 ms, loads in ~252 ms, and holds peak memory around 530–542 MiB.

Treat those figures as vendor benches on their pipeline — useful for sizing, not a substitute for your own domain audio.

How to Run It

Laptop / research path with NeMo (16 kHz mono):

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.from_pretrained(
    "qvac/TranscribePsy-250M-EU-sub"
).eval()
transcripts = model.transcribe(["audio_en.wav", "audio_de.wav"])
print([t.text for t in transcripts])

For the 25-language checkpoint, QVAC’s model card shows language-known (--lang en / --lang other) and gated auto-routing scripts after you download the repo folder. Phones should use the GGUF or Core ML builds through the qvac-transcribepsy / fabric-speech.cpp stack rather than full NeMo.

Where It Fits

If you need open, private, offline European ASR that fits a phone budget, TranscribePsy is a same-day drop with clear HF cards, Apache weights, and real-device timing. If you need live streaming captions or languages outside the EU set, wait for the streaming / broader-language work QVAC already flags as next — or stay on cloud ASR for those jobs. Specialist vocab (medical, legal) will likely want fine-tuning; QVAC says so up front.

Conclusion

TranscribePsy is a concrete open on-device ASR ship: 254M, Apache 2.0, 25 / 8 EU languages, NeMo + GGUF + Core ML, best-in-class WER in its size band on the suites they published, and phone numbers that make edge dictation feel real. Grab the HF collection, run a short multilingual clip on your target device, and decide whether the privacy-and-latency trade beats your current cloud STT.

—Anabel ♡