Fish Speech S2 Pro: local setup and long-form TTS

This tutorial covers a local setup for Fish Speech S2 Pro and the speaker's modified interface for long-form text-to-speech. The speaker describes the official web UI as suitable for short passages and says their fork can process scripts of 3,000 to 5,000 words. The demonstration uses roughly 1,000 words, with generation split into chunks. The modified interface defaults to 200 words per chunk and lets users choose a smaller size.

The installation walkthrough uses a Windows command prompt, Conda and CUDA 12.8. It also discusses Linux dependencies, including FFmpeg, and a CPU installation option. The speaker recommends about 24 GB of VRAM for GPU inference and describes the model files as roughly 11 GB. Setup includes downloading the S2 Pro checkpoints and Whisper Small, then supplying paths for the speech model, decoder and transcription model when launching the interface.

Whisper transcribes reference audio locally for the voice cloning workflow. The examples test accents and emotion tags before moving to a longer script. The speaker considers the cloned voice accurate but notes that missing pauses between paragraphs make the result sound rushed. The closing section covers BF16 precision, an FP16 half option, device selection and optional torch compile. It does not demonstrate a lower-memory configuration or measure a speed improvement.