Speech
QVAC TranscribePsy
TranscribePsy turns speech into text on the device itself, in 25 European languages on a laptop and 8 of them on a phone. On the Neural Engine of an iPhone 17 Pro Max, the 8-language build transcribes a 10-minute recording in 1.65 seconds.
Voice on device
Your words on screen before you lower the phone
Smart glasses, home robots and phones are moving to voice, and each of them first has to write down what you said. Accurate speech recognition models are large, so that step usually runs on a server, and it stops when the signal drops.
TranscribePsy is small enough to run inside a phone app, using about 530 MiB of memory at peak. It transcribes a spoken phrase in about a second on average across the six phones tested, faster than every other speech model compared, and the recording never leaves the phone.
What people use it for
Dictate notes
Speak a note or a message and get the text right after you stop.
Transcribe meetings
Turn a meeting or a lecture into text on the phone that recorded it.
Caption for accessibility
Give someone who is hard of hearing the text of each sentence a moment after it is said, with no connection needed.
Keep sensitive talks local
Write down a consultation, an interview or a source's account without the audio leaving the device.
At a glance
Specs
TranscribePsy comes as two models. TranscribePsy-250M-EU covers 25 European languages on a laptop. TranscribePsy-250M-EU-sub covers 8 of them with fewer errors, and also comes as a GGUF build for phone processors and a Core ML build for the iPhone Neural Engine.
- Size
- 254M parameters, in two models: EU and EU-sub
- Languages
- EU: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian · EU-sub: Dutch, English, French, German, Italian, Polish, Portuguese and Spanish
- Runs on
- Phones and laptops · measured on iPhone 17 Pro Max, iPhone 15, Galaxy S26 Ultra, Galaxy S25 Ultra, Pixel 10 and Pixel 9
- Speed
- 8-language build · a 10-minute recording in 1.65 seconds on the iPhone 17 Pro Max Neural Engine · 2.6 times faster than the next-fastest model tested on phone processors
- Memory
- About 530 to 540 MiB at peak · 8-bit build on a phone processor
- Start-up
- About a quarter of a second · 8-bit build on a phone processor
- Input and output
- 16 kHz mono audio in, text out · complete recordings and phrases · streaming transcription in progress
- Built on
- NVIDIA Parakeet-TDT 0.6B v3, a FastConformer model, pruned to 254M and distilled · NVIDIA NeMo toolkit
- License
- Apache 2.0, open weights
Results
Benchmarks
On three benchmark suites, TranscribePsy has the lowest word error rate of the similar-size models tested, from 242M to 370M parameters. On Multilingual LibriSpeech, averaged over its eight languages, TranscribePsy-250M-EU makes 6.58 word errors per 100 words, about half the error rate of every model its size, and it also beats four models three to four times larger on that benchmark.
All QVAC models on Hugging FaceWord error rate, Multilingual LibriSpeech
8 languages · lower is better
- TranscribePsy-250M-EU-sub5.87
- TranscribePsy-250M-EU6.58
- Whisper-small11.51
- OWSM v4 small11.84
- omniASR-CTC-300M-v211.89
Get the weights
Download TranscribePsy
Model
Size
Best for
Download
On Hugging Face
TranscribePsy-250M
TranscribePsy-250M-EU-sub-gguf
254M · At release
8 languages · phone processors, Android and iPhone
The GGUF build is 8-bit (Q8_0) and the Core ML build uses 6-bit weights. Both run through the qvac-transcribepsy engine. Download sizes marked At release are published with the models.
TranscribePsy is released under Apache 2.0 with open weights.
FAQ
Questions
Is TranscribePsy free?
Yes. The weights are open under Apache 2.0 and free to download from Hugging Face, along with the evaluation configs behind every published number.
Does it need an internet connection?
No. It runs on the phone or the laptop itself, so it keeps working with no signal and the audio stays on the device.
Which version should I pick?
Take TranscribePsy-250M-EU when you need any of its 25 languages, on a laptop. Take TranscribePsy-250M-EU-sub for its 8 languages, where it makes fewer errors. On a phone, use its GGUF build on the processor or its Core ML build on the iPhone Neural Engine. For the best performance and fastest inference speeds, run it on the parakeet engine of qvac-fabric-speech.cpp, which has been optimized for efficient execution and delivers the highest performance among the available runtime options.
Can it transcribe as I speak?
Not yet. This release transcribes each recording or phrase once it is complete. Streaming transcription is in progress.
What are its limits?
It covers 25 European languages on a laptop and 8 on a phone, and does not handle others. Fine-tune it when specialist vocabulary matters, and never use a transcript as the only input to a high-stakes automated decision.
Get started
Run AI on what you own
One install, JavaScript or Python. 10+ AI tasks. Private, offline and free on the hardware you already own.