Year: 2024
Type: dissertation
Abstract: Neural codec language models (NCLMs) are speech synthesizers that address the TTS task as a language modeling task rather than continous signal regression as in previous work. They showed an impressive generalization capability, surpassing previous state-of-the-art zero-shot TTS models by means of speaker similarity and naturalness. Although addressing speech synthesis as a language modeling task in part allows to train on large and diverse speech data crawled from the Internet, it also brings some issues common to those of large language models (LLMs) for text generation. While LLMs may generate outputs with made up facts or biased and toxic contents, neural codec language models suffers from synthesis robustness and expressivness. Reinforcement Learning from Human Feedback (RLHF) has emerged to tackle the issues of LLMs, by using human feedback to align the generated responses to the user preferences. Using RLHF helped LLMs to reduce the amount of generated toxic content and false facts. Motivated by the success of RLHF in the text generation domain, this work proposes to fine-tune NCLMs from feedback with reinforcement learning following the RLHF training pipeline. We conduct a series of experiments with VALL-E pretrained on LibriTTS, fine-tuning it to optimize different kind of feedback: intelligibility, naturalness, speaker similarity and waveform duration. Our results show that fine-tuning helped to increase the intelligibility of the model, showing a WER reduction up to 20.954\\%, but also to change the speech duration according to the reward signal. Finally, we delineate limitations of our experimental setup and propose practical mitigations, to be explored in future work.
Author: Pratesi, Lorenzo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7019341752