arXiv:2606.17537v1 Announce Type: cross Abstract: Non-autoregressive (NAR) decoding generates output tokens in parallel, making speech recognition faster than autoregressive decoding, which generates them sequentially from left to right. However, the recognition performance is degraded because NAR decoding cannot resolve uncertainty by conditioning on previously generated tokens. To address this issue, we propose a novel NAR decoding framework based on minimum Bayes' risk (MBR) decoding, termed NAR-MBR decoding, that maximizes the expected utility calculated from samples drawn from the output
The continuous drive for more efficient and faster AI processing, particularly in real-time applications like speech recognition, necessitates innovations that overcome current architectural limitations.
Improved speech recognition speed without significant accuracy degradation enhances user experience, enables new applications, and reduces computational overhead for large-scale deployments.
Speech recognition systems can now adopt architectures that offer faster processing with a new decoding framework that mitigates the traditional accuracy trade-off of non-autoregressive models.
- · AI product developers
- · Cloud providers
- · Speech-to-text service companies
- · Edge AI hardware manufacturers
Faster and more responsive voice interfaces become more ubiquitous across devices and services.
Reduced processing costs for speech AI could accelerate adoption in new, cost-sensitive markets.
This efficiency gain may free up compute resources, indirectly supporting other AI research and development areas.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL