Improving Speaker-Independent Lipreading with Domain-Adversarial Training
This addresses the challenge of speaker variability in lipreading systems, enabling more robust speech recognition from visual cues with limited adaptation data.
The paper tackled the problem of speaker-independent lipreading by using domain-adversarial training to adapt a lipreading system to new speakers with minimal untranscribed data, achieving a relative accuracy improvement of around 40% with only 15 to 20 seconds of target data.
We present a Lipreading system, i.e. a speech recognition system using only visual features, which uses domain-adversarial training for speaker independence. Domain-adversarial training is integrated into the optimization of a lipreader based on a stack of feedforward and LSTM (Long Short-Term Memory) recurrent neural networks, yielding an end-to-end trainable system which only requires a very small number of frames of untranscribed target data to substantially improve the recognition accuracy on the target speaker. On pairs of different source and target speakers, we achieve a relative accuracy improvement of around 40% with only 15 to 20 seconds of untranscribed target speech data. On multi-speaker training setups, the accuracy improvements are smaller but still substantial.