Mas Mollà, Gerard ; Sanchis, Albert ; Juan, Alfons Improving streaming ASR with foundation models using emission policies Inproceedings Proc. Interspeech 2026, pp. 4094–4098, Sidney (Australia), 2026. Abstract | Links | BibTeX | Tags: speech foundation models, streaming decoding @inproceedings{Mas2026,
title = {Improving streaming ASR with foundation models using emission policies},
author = {Mas Mollà, Gerard AND Sanchis, Albert AND Juan, Alfons},
doi = {10.21437/Interspeech.2026-3358},
year = {2026},
date = {2026-01-01},
booktitle = {Proc. Interspeech 2026},
pages = {4094--4098},
address = {Sidney (Australia)},
abstract = {Speech Foundation Models (SFMs) have achieved state-of-the-art results in offline ASR, yet their application in streaming environments remains limited due to the performance degradation caused by restricted context. In this work, we propose a model-agnostic, training-free streaming pipeline designed to wrap around ASR models in a "black box" manner. Our approach integrates three key components: a sliding window audio buffer for incremental data ingestion, token-level timestamp management to eliminate redundant output, and text-only emission policies to govern output timing. We systematically evaluate several emission policies, assessing their impact on the latency-accuracy trade-off across three English datasets. Results show that our pipeline enables timestamp-producing SFMs like Parakeet and Canary to achieve near-offline transcription quality in real-time environments without requiring access to internal model tensors.},
keywords = {speech foundation models, streaming decoding},
pubstate = {published},
tppubtype = {inproceedings}
}
Speech Foundation Models (SFMs) have achieved state-of-the-art results in offline ASR, yet their application in streaming environments remains limited due to the performance degradation caused by restricted context. In this work, we propose a model-agnostic, training-free streaming pipeline designed to wrap around ASR models in a "black box" manner. Our approach integrates three key components: a sliding window audio buffer for incremental data ingestion, token-level timestamp management to eliminate redundant output, and text-only emission policies to govern output timing. We systematically evaluate several emission policies, assessing their impact on the latency-accuracy trade-off across three English datasets. Results show that our pipeline enables timestamp-producing SFMs like Parakeet and Canary to achieve near-offline transcription quality in real-time environments without requiring access to internal model tensors. |
