Fine-Grained Speaking Style Control in Prompt-Based TTS Models

Researchers have introduced methods to achieve detailed control over speaking styles in prompt-based text-to-speech (TTS) models. This study tackles the shortcomings of existing models, which typically enforce a uniform style throughout an utterance, limiting applications that demand gradual style shifts over multiple utterances or within a single utterance. To facilitate style interpolation between utterances, the team calculates direction vectors from contrasting style prompts in the embedding space. Additionally, they recognize a significant attention bias toward initial tokens in autoregressive TTS decoders and propose KV-cache manipulation to counteract this issue, allowing for flexible style modifications within a single utterance.