Advancing end-to-end automatic speech recognition and beyond

The speech community is transitioning from hybrid modeling to end-to-end (E2E) modeling for automatic speech recognition (ASR). While E2E models achieved state-of-the-art results in most benchmarks in terms of ASR accuracy, there are lots of practical factors that affect the production model deployment decision, including low-latency streaming, leveraging text-only data, and handling overlapped speech etc. Without providing excellent solutions to all these factors, it is hard for E2E models to be widely commercialized.

In this talk, I will overview the recent advances in E2E models with the focus on technologies addressing those challenges from the perspective of industry. To design a high-accuracy low-latency E2E model, a masking strategy was introduced into Transformer Transducer. I will discuss technologies which can leverage text-only data for general model training via pretraining and adaptation to a new domain via augmentation and factorization. Then, I will extend E2E modeling for streaming multi-talker ASR. I will also show how we go beyond ASR by extending the learning in E2E ASR into a new area like speech translation and build high-quality E2E speech translation models even without any human labeled speech translation data. Finally, I will conclude the talk with some new research opportunities we may work on.

日期：
演讲者：
Jinyu Li
所属机构：
International Symposium on Chinese Spoken Language Processing

接下来观看