Persistent Delivery Optimization for Streaming Speech-to-Text Translation with Revisions
En palabras de los autores
Revision-capable streaming speech-to-text translation (S2TT) can correct earlier drafts, but process rewards based on visible text may credit content later withdrawn. Persistent Delivery Optimization (PDO) assigns intermediate reward only to content that survives revisions while scoring final quality separately. With 7.49 h of task-specific FLEURS adaptation, PDO achieves the best BLEU on four of five directions and higher COMET than every external streaming baseline in all five directions. Relative to its History-SFT initialization, PDO reduces mean/P90 finalization-aware latency by 10.8%/11.3% and normalized erasure by 15.8%, while emitting at the first permitted 2-s update and improving macro BLEU. Zero-shot evaluation on Europarl-ST and CoVoST 2 confirms that these gains are not confined to the FLEURS training domain.
Apareció: miércoles, 23 de septiembre. arXiv. Preprint, todavía sin revisión por pares.
Comentario de los autores: 5 pages, 1 figure, 4 tables. Training code and model weights are available at https://github.com/ggiggit/PDO_S2TT. Submitted to ICASSP 2027