Audio-Native Speech Recognition with a Frozen Discrete-Diffusion Language Model
Harsha Vardhan Khurdula, Abhinav Kumar Singh, Yoeven D Khemlani, Vineet Agarwal
Read on arXiv →
Key claim
Discrete diffusion model improves speech recognition efficiency.
In plain English
Automatic speech recognition typically relies on autoregressive models that generate one token at a time, which can be slow and inefficient. Current methods often struggle to effectively ground audio features in the transcription process. This paper introduces a discrete diffusion model that refines entire transcripts in parallel, significantly improving transcription speed and accuracy. Builders might care about this approach as it offers a new way to handle speech recognition tasks more efficiently, especially in multilingual contexts.
Introduces a novel approach to speech recognition using discrete diffusion models.
Demonstrates solid performance on a standard benchmark with a clear methodology.
Deep reliability assessment
The methodology supports the use of a discrete diffusion language model for speech transcription, but the claim of achieving competitive quality with autoregressive models may be overclaimed without direct comparison.
Reproducibility
No open source code or dataset is mentioned in the paper.
Key figure
The key architectural diagram likely illustrates the integration of a frozen Whisper encoder with a lightweight projector and low-rank adapters into the DiffusionGemma model.