Interfaze released an open-source diffusion ASR model for speech recognition in six languages

Interfaze published diffusion-gemma-asr-small in open access on July 2, 2026 — a multilingual automatic speech recognition model that transcribes audio through diffusion rather than autoregression. An adapter weighing about 42 million parameters connects to frozen DiffusionGemma weights from Google and covers six languages without splitting into language-specific branches.

How does diffusion speech recognition work?

Most public ASR systems are autoregressive: the model generates transcription sequentially, token by token, and execution time grows proportionally to the length of the recording. Whisper from OpenAI, MMS and Seamless from Meta, Distil-Whisper — they all work exactly this way.

diffusion-gemma-asr-small is organized differently. At its core is DiffusionGemma — Google's language model trained with the diffusion method: instead of sequential generation, it recovers text from a noisy state in parallel through several denoising steps.

Interfaze added an audio adapter to this foundation: it teaches the model to correlate the audio signal with the desired text, while the weights of DiffusionGemma remain frozen and are not updated during adapter training.

The practical consequence: the computational cost of transcription is determined by the specified number of denoising steps, not the length of audio. The developer themselves chooses the balance between speed and quality — more steps means higher accuracy, but longer time.

What's inside the model?

• Release date in open access — July 2, 2026
• Size of audio adapter — about 42 million parameters
• Base model — DiffusionGemma from Google (weights frozen, not updated)
• Supported languages — six (specific list not disclosed in the announcement)
• Inference cost depends on the number of denoising steps, not on the length of the recording

One adapter for six languages — 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, and the adapter only needs to teach it to accept audio — without duplicating weights for each language.

For comparison: the full version of Whisper large-v3 from OpenAI contains 1.5 billion parameters — Interfaze's adapter is approximately 35 times more compact in terms of retrainable weights.

What does this mean for ASR developers?

The diffusion approach in speech recognition is rare in public space. 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, this is proof of concept and a starting point: open weights allow exploring the architecture and trying to adapt it to related tasks — multilingual ASR, code-switching, accent-aware speech recognition.

For practical application, data is needed, which the announcement does not disclose: specific languages, accuracy on standard benchmarks (WER on LibriSpeech, Common Voice, Fleurs), real inference speed compared to Whisper. Without these numbers, it is difficult to assess the model's competitiveness in production scenarios.

What this means

Interfaze opened a practically unexplored direction — diffusion ASR with retraining only a light adapter on top of an existing language model. If the approach confirms competitive quality on benchmarks, it will offer an interesting alternative to autoregressive systems: predictable latency, flexible quality adjustment through the number of steps, and a compact weight budget.

*Meta is recognized as an extremist organization and is banned in Russia.