Interfaze Ships diffusion-gemma-asr-small, an Open-Source Diffusion ASR Model Transcribing Six Languages via DiffusionGemma’s Parallel Denoising Decoder

An open-source speech-to-text model called diffusion-gemma-asr-small converts audio to text using a diffusion decoder rather than the more common autoregressive approach, meaning it refines all output tokens in parallel instead of generating them one at a time. The model handles six languages through a single adapter of about 42 million parameters attached to a much larger frozen backbone, making it efficient for multilingual transcription without loading separate models per language. Unlike autoregressive models that generate output length-dependent costs, this diffusion approach scales cost by denoising steps instead, making it well-suited for batch transcription where longer and shorter clips require similar processing time. While it performs competitively against other diffusion-based speech recognition systems, it trails the autoregressive Whisper model on benchmarks, a gap the research team attributes to training data rather than architectural differences.