On the Problem of Text-To-Speech Model Selection for Synthetic Data Generation in Automatic Speech Recognition
This addresses the challenge for ASR researchers in choosing TTS systems for synthetic data generation, but it is incremental as it compares existing architectures without introducing a new method.
The study tackled the problem of selecting text-to-speech (TTS) models for generating synthetic data in automatic speech recognition (ASR), finding that auto-regressive decoding outperforms non-autoregressive decoding and that metrics like NISQA MOS do not clearly relate to ASR performance.
The rapid development of neural text-to-speech (TTS) systems enabled its usage in other areas of natural language processing such as automatic speech recognition (ASR) or spoken language translation (SLT). Due to the large number of different TTS architectures and their extensions, selecting which TTS systems to use for synthetic data creation is not an easy task. We use the comparison of five different TTS decoder architectures in the scope of synthetic data generation to show the impact on CTC-based speech recognition training. We compare the recognition results to computable metrics like NISQA MOS and intelligibility, finding that there are no clear relations to the ASR performance. We also observe that for data generation auto-regressive decoding performs better than non-autoregressive decoding, and propose an approach to quantify TTS generalization capabilities.