Interfaze Ships diffusion-gemma-asr-small, an Open-Source Diffusion ASR Model Transcribing Six Languages via DiffusionGemma’s Parallel Denoising Decoder
Interfaze just shipped a diffusion-based ASR model that does what autoregressive can't: transcribe six languages with a single 42M adapter.
Why it matters
Open-source diffusion ASR challenges the autoregressive paradigm for speech recognition, offering a novel inference approach (denoising steps vs. transcript length) that could reshape how companies think about transcription cost structures and multilingual deployment.
The key facts
10 to knowInterfaze open-sourced diffusion-gemma-asr-small
Uses diffusion-based decoding instead of autoregressive generation
~42M-parameter adapter on frozen Google DiffusionGemma
Single adapter covers six languages
Transcription cost determined by denoising steps, not transcript length
Multilingual capability with unified model architecture
42M-parameter adapter architecture
Covers 6 languages with single adapter
Built on Google's frozen DiffusionGemma base
Cost scales with denoising steps, not transcript length
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Interfaze open-sourced diffusion-gemma-asr-small, a multilingual ASR model that transcribes via diffusion, not autoregression. It adds audio to Google's frozen DiffusionGemma using a ~42M-parameter adapter. One adapter covers six languages, with transcription cost set by denoising steps, not…