On the Audio-visual Synchronization for Lip-to-Speech Synthesis
This addresses data quality and model synchronization issues in lip-to-speech synthesis, which is important for applications like assistive technologies, but it is incremental as it builds on existing methods by adding synchronization mechanisms.
The paper tackles the problem of audio-visual asynchrony in lip-to-speech synthesis datasets, which causes models to generate out-of-sync speech, and proposes a synchronized model with an automatic mechanism that improves performance, showing benefits over state-of-the-art methods on both conventional and time-aligned metrics.
Most lip-to-speech (LTS) synthesis models are trained and evaluated under the assumption that the audio-video pairs in the dataset are perfectly synchronized. In this work, we show that the commonly used audio-visual datasets, such as GRID, TCD-TIMIT, and Lip2Wav, can have data asynchrony issues. Training lip-to-speech with such datasets may further cause the model asynchrony issue -- that is, the generated speech and the input video are out of sync. To address these asynchrony issues, we propose a synchronized lip-to-speech (SLTS) model with an automatic synchronization mechanism (ASM) to correct data asynchrony and penalize model asynchrony. We further demonstrate the limitation of the commonly adopted evaluation metrics for LTS with asynchronous test data and introduce an audio alignment frontend before the metrics sensitive to time alignment for better evaluation. We compare our method with state-of-the-art approaches on conventional and time-aligned metrics to show the benefits of synchronization training.