Tavern Talk

A custom web interface for text-to-speech using Fish Speech 1.5. Built with React + TypeScript, runs entirely in Docker with GPU acceleration.

Requirements

• Docker with NVIDIA Container Toolkit
• NVIDIA GPU (~2GB VRAM)

Quick Start

git clone <this-repo>
cd tavern-talk
docker compose --profile dev up

On first run this downloads the Fish Speech 1.5 model checkpoints (~1.4GB). Once you see Vite ready and File API listening in the logs, open http://localhost:3000.

Features

• Voice Synthesis -- type text, adjust parameters, generate audio with one click
• Full API control -- all Fish Speech parameters exposed (temperature, top_p, repetition penalty, seed, chunk length, max tokens, streaming, normalize, memory cache)
• Reproducible output -- every generation gets a seed (auto-generated if not set), shown in the settings snapshot so you can recreate any result
• Presets -- save parameter settings as named presets, load them to instantly apply
• Voice Profiles -- upload reference audio to clone voices, with auto-transcription
• Generation History -- browse past generations, replay audio, re-apply settings
• Settings -- configure defaults, check API health
• Hot-reload development -- edit React components, see changes instantly

Architecture

Browser :3000  -->  Vite dev server (React SPA)
       /api/*  -->  Fish Speech API (port 8080, internal)
     /files/*  -->  Express file API (port 3001, internal)

Three Docker services:

Service	Purpose
api	Fish Speech 1.5 TTS engine (GPU)
frontend-dev	Vite + Express file API (development)
frontend-prod	nginx + Express file API (production)

Development

# Start with hot-reload
docker compose --profile dev up

# Rebuild after dependency changes
docker compose --profile dev up --build

Edit files in frontend/src/ -- Vite hot-reloads automatically. The source directory is volume-mounted into the container.

Production

docker compose --profile prod up --build

Builds the React app, serves static files via nginx on port 3000.

Stopping

docker compose --profile dev down

Project Structure

Path	Description
frontend/	React + TypeScript app (Vite)
frontend/src/pages/	Synthesis, Voices, History, Settings pages
frontend/src/components/	Shared UI components
frontend/src/api/	Typed API clients (TTS, ASR, file operations)
frontend/src/stores/	Zustand state management
frontend/server/	Express file API sidecar
docker-compose.yml	Service orchestration with dev/prod profiles
checkpoints/	Model weights, auto-downloaded on first run (git-ignored)
voices/	Saved voice profiles (git-ignored)
output/	Generation history and settings (git-ignored)
samples/	Design docs and mockups

CLI Generation

The generate.sh script can generate audio directly from the command line. It requires the API to be running (start with docker compose --profile dev up):

./generate.sh "Welcome to the tavern, adventurer."
./generate.sh "Hello world" output/hello.wav

Tech Stack

• Frontend: React 19, TypeScript, Vite, React Router, Zustand, Radix UI Tooltip, Tabler Icons
• File API: Express, multer
• TTS Engine: Fish Speech 1.5 (fishaudio/fish-speech:v1.5.1)
• Infrastructure: Docker, nginx

Notes

• Fish Speech 1.5 is licensed CC BY-NC-SA 4.0 (non-commercial use)
• All GPU devices are passed through to the API container
• Voice profiles and generation history persist in mounted volumes (voices/, output/)