Interfaze releases open diffusion-based ASR model covering six languages

On July 2, 2026, Interfaze released diffusion-gemma-asr-small publicly. It is a multilingual automatic speech recognition model that transcribes audio through diffusion rather than autoregression. The adapter, at about 42 million parameters, plugs into the frozen weights of Google's DiffusionGemma and covers six languages without splitting into language-specific branches.

How does diffusion-based speech recognition work?

Most public ASR systems are autoregressive: the model generates the transcript sequentially, token by token, and runtime grows in proportion to the length of the recording. OpenAI's Whisper, Meta's MMS and Seamless, and Distil-Whisper all work this way.

diffusion-gemma-asr-small is built differently. It rests on DiffusionGemma, a Google language model trained with diffusion: instead of generating text sequentially, it reconstructs it from a noised state in parallel, over several denoising steps. Interfaze added an audio adapter on top. The adapter teaches the model to map the audio signal to the right text, while DiffusionGemma's weights stay frozen and are not updated during adapter training.

The practical consequence: the compute cost of transcription is set by the chosen number of denoising steps, not by audio length. Developers choose the trade-off between speed and quality themselves: more steps mean higher accuracy but longer processing time.

What is inside the model?

• Open-source release date: July 2, 2026
• Audio adapter size: about 42 million parameters
• Base model: Google's DiffusionGemma (weights frozen, not updated)
• Supported languages: six (the announcement does not list them)
• Inference cost depends on the number of denoising steps, not on recording length

One adapter for six languages is the key architectural bet. Most multilingual models use either language-specific heads or a separate checkpoint for each language. Here, DiffusionGemma already contains multilingual text representations, so the adapter only has to teach it to accept audio, with no duplicated weights per language.

For comparison, OpenAI's full Whisper large-v3 has 1.5 billion parameters, so the Interfaze adapter is roughly 35 times smaller in trainable weights.

What does this mean for ASR developers?

Diffusion is rare in public speech recognition work. Most open ASR systems are built on autoregression or CTC decoding. diffusion-gemma-asr-small is the first public example of reusing a diffusion language model as an ASR backend through a minimal adapter.

For the research community, it is a proof of concept and a starting point: the open weights make it possible to study the architecture and try adapting it to related tasks such as multilingual ASR, code-switching and accented speech recognition.

Practical use requires data the announcement does not disclose: the specific languages, accuracy on standard benchmarks (WER on LibriSpeech, Common Voice, Fleurs), and real inference speed compared with Whisper. Without these figures, it is hard for now to judge how competitive the model is in production scenarios.

What this means

Interfaze has opened up a largely unexplored direction: diffusion ASR that fine-tunes only a lightweight adapter on top of a ready-made language model. If the approach shows competitive quality on benchmarks, it would offer an interesting alternative to autoregressive systems: predictable latency, quality that can be tuned through the number of steps, and a compact weight budget.

*Meta has been designated an extremist organization and is banned in Russia.