NAR-MBR decoding boosts speech recognition speed without accuracy loss
New method decodes speech in parallel, outpacing autoregressive models on 4 datasets.
Deep Dive
A new non-autoregressive decoding framework, NAR-MBR, maximizes expected utility from multiple samples drawn from the NAR model's output probability. By leveraging NAR models to efficiently obtain samples in a single forward pass, it outperforms previous NAR decoding and runs faster than autoregressive decoding, as demonstrated on LibriSpeech, Switchboard, AMI, and web presentation corpus.
Key Points
- Proposes NAR-MBR decoding that applies minimum Bayes' risk to non-autoregressive speech recognition models.
- Leverages parallel token generation to sample multiple outputs in a single forward pass, improving accuracy without added latency.
- Outperforms previous NAR decoding and runs faster than autoregressive decoding across four benchmark datasets (LibriSpeech, Switchboard, AMI, web presentation corpus).
Why It Matters
Faster, more accurate speech recognition enables real-time transcription and voice interfaces without hardware upgrades.