SillyTavern PocketTTS WebSocket

TTS extension for pocket-tts-openapi.

Features

• Sentence-level streaming — audio plays during generation
• Server-reported audio duration per sentence
• Persistent WebSocket connection (single /v1/audio/stream endpoint)
• Player bar with seek, volume, speed, highlight, and playlist controls
• Auto-discovered voice cloning (.wav voices from server)
• CPU/GPU model selection
• Auto-cleanup of played tracks

Requirements

• pocket-tts-openapi server (default localhost:8005)
• SillyTavern

Architecture

Text Processing

Text is split on sentence boundaries (.!?…). Short sentences (<20 chars) are merged server-side. Each sentence becomes a track in a flat queue — requests go out in play order.

WebSocket

Single persistent connection to /v1/audio/stream. Requests are fire-and-forget with sequential IDs. Audio chunks arrive independently and are matched to requests by request_id on done message.

Highlighting

Text highlighting uses substring search against rendered DOM with Unicode normalization for quotes and ellipsis characters.

Settings

Setting	Default	Description
Server Endpoint	http://localhost:8005	Server URL
Model	english-cpu	Dynamically populated from server (e.g., english-cpu, french_24l-gpu)
Audio Format	mp3	mp3, wav, opus, flac, aac
Speed	1.0	0.5–2.0
Temperature	1.0	0.0–2.0
Top P	1.0	0.0–1.0

Voice mapping: SillyTavern's TTS Voice Map section.

GPU VRAM Overhead (CUDA Context)

When using a -gpu model variant, you may notice a lingering ~170 MB of VRAM usage even after TTS generation finishes and you switch back to a -cpu model. This is not a memory leak.

Whenever a PyTorch application initializes the GPU for the first time, the NVIDIA driver creates a "CUDA context" for that Python process. This context loads essential GPU libraries (like cuBLAS/cuDNN handles) and driver state into VRAM. Depending on your driver version and GPU architecture, this base footprint is typically between 150 MB and 300 MB.

This overhead is a hard architectural limitation of PyTorch and CUDA. Once a Python process creates a CUDA context, that memory is locked to the process until the backend server itself is shut down. However, the heavy components (the 1GB+ TTS model weights) are successfully freed when switching to a -cpu model in the extension UI, leaving the vast majority of your VRAM completely free for your LLMs and other tasks.

Feedback

Join my Discord: https://discord.gg/2tJcWeMjFQ Or find me on the official SillyTavern Discord server.

Support me: Patreon