Recurrent Neural Network Transducer for Audio-Visual Speech Recognition
About
This work presents a large-scale audio-visual speech recognition system based on a recurrent neural network transducer (RNN-T) architecture. To support the development of such a system, we built a large audio-visual (A/V) dataset of segmented utterances extracted from YouTube public videos, leading to 31k hours of audio-visual training content. The performance of an audio-only, visual-only, and audio-visual system are compared on two large-vocabulary test sets: a set of utterance segments from public YouTube videos called YTDEV18 and the publicly available LRS3-TED set. To highlight the contribution of the visual modality, we also evaluated the performance of our system on the YTDEV18 set artificially corrupted with background noise and overlapping speech. To the best of our knowledge, our system significantly improves the state-of-the-art on the LRS3-TED set.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Visual Speech Recognition | LRS3 (test) | WER4.5 | 159 | |
| Visual Speech Recognition | LRS3 High-Resource, 433h labelled v1 (test) | WER0.045 | 80 | |
| Audio-Visual Speech Recognition | LRS3 clean (test) | WER4.5 | 70 | |
| Visual Speech Recognition | LRS3 | WER0.045 | 59 | |
| Automatic Speech Recognition | LRS3 (test) | -- | 46 | |
| Speech Recognition | LRS3-TED | WER33.6 | 25 | |
| Visual Speech Recognition | LRS3 low-resource (test) | WER33.6 | 20 | |
| Lip-reading | LRS3 1.0 (test) | WER33.6 | 19 | |
| Automatic Speech Recognition | LRS3 433-hour labeled (test) | WER (%)4.8 | 19 | |
| Automatic Speech Recognition | LRS3 low-resource (test) | WER0.048 | 18 |