What is pilottts?

PilotTTS is a Python text-to-speech system that converts written text into spoken audio. It was built by the voice team at Amap (the Chinese mapping and navigation service) and released alongside a research paper. The model weights are available on HuggingFace and ModelScope. The system has two modes. The base model does zero-shot voice cloning: you provide a short audio clip of any speaker, and the model generates new speech in that person's voice from text you supply. The instruct model adds three layers of control on top of voice cloning. First, you can specify one of twelve emotions such as happy, sad, angry, or serious, and the output adjusts its tone accordingly. Second, you can insert paralinguistic markers directly into the text, for example placing a laugh tag at a point in a sentence to add a natural-sounding laugh. The supported sounds include laughter, breathing, crying, and coughing. Third, the instruct model supports fourteen Chinese regional dialects, including Shanghainese, Sichuanese, Minnan, and several others, so you can generate speech that sounds like it comes from a specific region. The project is built from entirely open-source components. It uses a small language model called Qwen3-0.6B to process text, a Meta audio feature extractor called w2v-bert-2.0, and a vocoder from a project called CosyVoice3 that converts the model's output into actual audio. Setup involves creating a Python environment, installing dependencies, and downloading the pretrained model files from HuggingFace or ModelScope. You can run inference from a Python script, from the command line, or through a browser-based interface launched with a single command. The browser interface is built with Gradio. A demo script runs examples of all inference modes at once so you can verify the installation is working. The README is written in English with a Chinese version linked separately. The full README is longer than what was shown.