The `Text Processing Module` is a critical subsystem responsible for preparing raw input text for the speech synthesis pipeline. It ensures that the text is linguistically sound and converted into a numerical format suitable for machine learning models. This module comprises two fundamental components: the `Text Normalizer` and the `Text Tokenizer`.

Components

Text Normalizer

This component, implemented by the `ChatTTS.norm.Normalizer` class, is responsible for the linguistic pre-processing of raw input text. Its primary functions include standardizing punctuation (e.g., converting various symbols to a consistent set), applying character mappings (e.g., half-width to full-width conversions), performing homophone replacement to resolve ambiguities, and removing invalid characters. It also includes language detection to apply language-specific normalizations. The goal is to ensure the text is clean, consistent, and correctly interpreted before tokenization, which is crucial for accurate pronunciation by the speech synthesis model.

Text Tokenizer

This component, implemented by the `ChatTTS.model.tokenizer.Tokenizer` class, serves as the interface between human-readable text and the numerical representations required by the machine learning models. It takes the normalized text and converts it into numerical token IDs and corresponding attention masks using a pre-trained `BertTokenizerFast`. Conversely, it can also decode sequences of token IDs back into human-readable text. It handles special tokens relevant to the ChatTTS model, such as `[spk_emb]` and `[break_0]`.