Researchers demonstrate that a discrete diffusion language model can transcribe speech by refining a whole transcript in parallel over a small number of denoising steps, rather than using autoregressive decoders.
- The approach uses an audio-native interface for DiffusionGemma, a 26B mixture-of-experts model generating text via uniform, random-token discrete diffusion.
- A frozen Whisper encoder supplies acoustic features to a lightweight projector, with low-rank adapters allowing the frozen backbone to attend to the new modality.
- Only about 42M parameters are trained (0.16 percent of the backbone), using a connectionist temporal classification loss through the frozen output head to ground audio.
- The model achieves 6.6 percent word error rate on LibriSpeech test-clean and transcribes in roughly eight parallel steps regardless of utterance length.
The resulting system uses a single adapter trained on six languages, evaluated here on English, Hindi, and Mandarin, offering a parallel alternative to sequential speech recognition.