aiola / drax-v1

Drax: Speech Recognition with Discrete Flow Matching

Model Overview

The Drax model family provides speech recognition models based on discrete flow matching. The drax-v1 model supports eight languages: English, Spanish, French, Portuguese, German, Italian, Japanese and Chinese. It is an encoder-decoder model consists of a Whisper-large-v3 encoder, and a DiT based decoder, with a total of ~1.2B parameters.

More details on usage in our GitHub repo, https://github.com/aiola-lab/drax and our paper .

Usage

See https://github.com/aiola-lab/drax for installation instructions.

from drax import Transcriber

asr = Transcriber(model_path="aiola-labs/drax-v1")
result = asr.transcribe("/path/to/audio.wav", language="en")
print(result[0].transcript)

Control sampling steps, temperature etc.

from drax import Transcriber

asr = Transcriber(model_path="aiola-labs/drax-v1")
result = asr.transcribe("/path/to/audio.wav", language="en", sampling_steps=32, temperature=1e-2)
print(result[0].transcript)

Batch inference:

from drax import Transcriber

asr = Transcriber(model_path="aiola-labs/drax-v1")
audio_paths = ["/path/to/audio1.wav", "/path/to/audio2.wav"]
languages = ["en", "de"]
result = asr.transcribe(audio_paths, language=languages)
print(result.transcript)

Citation

@article{navon2025drax,
  title={Drax: Speech Recognition with Discrete Flow Matching},
  author={Navon, Aviv and Shamsian, Aviv and Glazer, Neta and Segal-Feldman, Yael and Hetz, Gill and Keshet, Joseph and Fetaya, Ethan},
  journal={arXiv preprint arXiv:2510.04162},
  year={2025}
}