ChatTTS

Generates natural speech from text using a GPT-based architecture specifically trained for conversational dialogue, with fine-grained control over prosodic features including laughter, pauses, and interjections. The system uses a two-stage pipeline: optional GPT-based text refinement that injects prosody markers into the input, followed by discrete audio token generation via a transformer-based audio codec. This approach enables expressive, contextually-aware speech synthesis rather than flat, robotic output typical of generic TTS systems.

Refines raw input text by running it through a fine-tuned GPT model that adds prosody markers (e.g., [laugh], [pause], [breath]) and improves phrasing for natural speech synthesis. The GPT model operates on discrete tokens and outputs enriched text that guides the downstream audio codec toward more expressive speech. This refinement is optional and can be disabled via skip_refine_text=True for latency-critical applications, but enabling it significantly improves speech naturalness by making the model aware of conversational context.

Implements GPU acceleration for all computationally expensive stages (text refinement, token generation, spectrogram decoding, vocoding) using PyTorch and CUDA, enabling real-time or near-real-time synthesis on modern GPUs. The system automatically detects GPU availability and moves models to GPU memory, with fallback to CPU inference if needed. GPU optimization includes batch processing, kernel fusion, and memory management to maximize throughput and minimize latency.

Exports trained models to ONNX (Open Neural Network Exchange) format, enabling deployment on diverse platforms and runtimes without PyTorch dependency. The system supports exporting the GPT model, DVAE decoder, and Vocos vocoder to ONNX, enabling inference on CPU-only servers, edge devices, or specialized hardware (e.g., NVIDIA Triton, ONNX Runtime). ONNX export includes quantization and optimization options for reducing model size and inference latency.

Supports synthesis for both English and Chinese languages with language-specific text normalization, tokenization, and prosody handling. The system automatically detects input language or allows explicit language specification, routing text through appropriate language-specific pipelines. Language support includes both Simplified and Traditional Chinese, with separate models and tokenizers for each language to ensure accurate pronunciation and prosody.

Provides a web-based user interface for interactive text-to-speech synthesis, speaker management, and parameter tuning without requiring programming knowledge. The web interface enables users to input text, select or generate speakers, adjust synthesis parameters, and listen to generated audio in real-time. The interface is built with modern web technologies and communicates with the backend Chat class via HTTP API, enabling easy deployment and sharing.

Provides a command-line interface (CLI) for batch synthesis, enabling users to synthesize multiple utterances from text files or command-line arguments without writing Python code. The CLI supports common options like input/output paths, speaker selection, sample rate, and refinement control, making it suitable for scripting and automation. The CLI is built on top of the Chat class and exposes its core functionality through command-line arguments.

Generates sequences of discrete audio tokens (codes) from refined text and speaker embeddings using a transformer-based audio codec. The system encodes speaker characteristics (voice identity, timbre, pitch range) as continuous embeddings that condition the token generation process, enabling voice cloning and speaker variation without retraining the model. Audio tokens are discrete (typically 1024-4096 vocabulary size) rather than continuous, making them more stable and enabling better control over audio quality and speaker consistency.

Extracts speaker characteristics (voice identity, timbre, pitch range) from a reference audio sample and encodes them as a continuous embedding vector that can be used to condition subsequent speech synthesis. The system uses the DVAE encoder to process the reference audio and extract speaker-specific features, enabling voice cloning without explicit speaker labels or manual parameter tuning. This embedding can then be reused across multiple synthesis calls to maintain speaker consistency.

Generates random speaker embeddings from a learned distribution, enabling diverse voice synthesis without reference audio or manual speaker specification. The system samples from the speaker embedding space (typically a Gaussian or learned distribution) to create novel speaker identities that are compatible with the synthesis model. This allows applications to generate speech with varied voices without requiring pre-recorded reference samples or explicit speaker parameters.

Cleans and standardizes input text before synthesis, handling language-specific features such as homophone replacement, number-to-word conversion, and punctuation normalization. The Normalizer component processes text to ensure consistent input to downstream models, handling edge cases like abbreviations, special characters, and language-specific conventions (e.g., Chinese number formatting). This preprocessing step is transparent to users but critical for robust synthesis across diverse input text.

Decodes discrete audio tokens into mel spectrograms using a DVAE (Discrete Variational Autoencoder) decoder, converting the compact token representation into a continuous acoustic representation suitable for vocoding. The DVAE decoder maps from discrete token space to continuous spectrogram space, enabling the separation of content (tokens) from acoustic details (spectrogram). This intermediate representation allows for flexible audio processing and quality control before final waveform generation.

Converts mel spectrograms into high-quality audio waveforms using Vocos, a neural vocoder trained on large-scale speech data. Vocos operates on mel spectrograms and generates raw waveforms at the target sample rate (16kHz or 24kHz), enabling fast, high-quality audio synthesis without traditional signal processing. The vocoder is a separate component that can be swapped or fine-tuned independently, providing flexibility for quality tuning or domain adaptation.

Processes multiple text utterances in a single inference call, enabling efficient batch synthesis with shared model state and optimized GPU utilization. The system batches text normalization, refinement, token generation, and decoding steps, reducing per-utterance overhead and enabling faster throughput for multi-utterance synthesis. Batch processing is transparent to users — the infer() method handles batching automatically based on input type (list of strings).

Provides fine-grained control over the inference pipeline through configuration parameters, including the ability to skip text refinement for latency-critical applications, control sample rates, and adjust decoding strategies. The Chat class exposes parameters like skip_refine_text, sample_rate, and decoder selection, enabling users to trade off between quality and latency. Configuration is managed through a central Config object that propagates settings through all pipeline stages.