CONF Thorbecke_EMNLP_2024/IDIAP Fast Streaming Transducer ASR Prototyping via Knowledge Distillation with Whisper Thorbecke, Iuliia Zuluaga-Gomez, Juan Villatoro-Tello, Esaú Kumar, Shashi Rangappa, Pradeep Burdisso, Sergio Motlicek, Petr S, Karthik Pandia D Ganapathiraju, Aravind pseudo-labelling shallow fusion streaming transducer EXTERNAL https://publications.idiap.ch/attachments/papers/2024/Thorbecke_EMNLP_2024.pdf PUBLIC https://publications.idiap.ch/index.php/publications/showcite/Iuliia_Idiap-RR-10-2024 Related documents Findings of the Association for Computational Linguistics: EMNLP 2024 2024 Association for Computational Linguistics (ACL) Miami, Florida, USA 16747–16762 https://aclanthology.org/2024.findings-emnlp.976/ URL 10.18653/v1/2024.findings-emnlp.976 doi The training of automatic speech recognition (ASR) with little to no supervised data remains an open question. In this work, we demonstrate that streaming Transformer-Transducer (TT) models can be trained from scratch in consumer and accessible GPUs in their entirety with pseudo-labeled (PL) speech from foundational speech models (FSM). This allows training a robust ASR model just in one stage and does not require large data and computational budget compared to the two-step scenario with pre-training and fine-tuning. We perform a comprehensive ablation on different aspects of PL-based streaming TT models such as the impact of (1) shallow fusion of n-gram LMs, (2) contextual biasing with named entities, (3) chunk-wise decoding for low-latency streaming applications, and (4) TT overall performance as the function of the FSM size. Our results demonstrate that TT can be trained from scratch without supervised data, even with very noisy PLs. We validate the proposed framework on 6 languages from CommonVoice and propose multiple heuristics to filter out hallucinated PLs. REPORT Iuliia_Idiap-RR-10-2024/IDIAP Fast Streaming Transducer ASR Prototyping via Knowledge Distillation with Whisper Iuliia, Thorbecke Zuluaga-Gomez, Juan Villatoro-Tello, Esaú Kumar, Shashi Rangappa, Pradeep Burdisso, Sergio Motlicek, Petr S, Karthik Pandia D Ganapathiraju, Aravind EXTERNAL https://publications.idiap.ch/attachments/reports/2024/Iuliia_Idiap-RR-10-2024.pdf PUBLIC Idiap-RR-10-2024 2024 Idiap October 2024 accepted to EMNLP The training of automatic speech recognition (ASR) with little to no supervised data remains an open question. In this work, we demonstrate that streaming Transformer-Transducer (TT) models can be trained from scratch in consumer and accessible GPUs in their entirety with pseudo-labeled (PL) speech from foundational speech models (FSM). This allows training a robust ASR model just in one stage and does not require large data and computational budget compared to the two-step scenario with pre-training and fine-tuning. We perform a comprehensive ablation on different aspects of PL-based streaming TT models such as the impact of (1) shallow fusion of n-gram LMs, (2) contextual biasing with named entities, (3) chunk-wise decoding for low-latency streaming applications, and (4) TT overall performance as the function of the FSM size. Our results demonstrate that TT can be trained from scratch without supervised data, even with very noisy PLs. We validate the proposed framework on 6 languages from CommonVoice and propose multiple heuristics to filter out hallucinated PLs.