VoiceCraft

Convert e-books (FB2, EPUB, TXT) to MP3 audiobooks using various Text-to-Speech technologies.

Features

• Six TTS providers - RHVoice, Piper, Silero, Coqui XTTS-v2, Bark Small and ElevenLabs
• 60+ voices - Multiple voices in Russian and English
• Format support - FB2, EPUB, TXT
• Speed control - from 0.5x to 2.0x
• Multiple themes - Light/dark/system theme support
• Multi-language UI - English and Russian interface
• Auto-splitting - large books split into parts
• GPU acceleration - NVIDIA CUDA and AMD DirectML support for faster generation

TTS Providers

1. RHVoice

Speed: Fast | Quality: Good | Offline

Lightweight offline engine based on Windows SAPI with minimal installation size (~15 MB per voice). Provides instant speech generation with very low CPU usage, making it perfect for converting large books quickly.

Russian voices:

• Aleksandr (Male)
• Irina, Anna, Elena (Female)

English voices:

• Bdl, Alan (Male)
• Slt, Clb (Female)

Download: RHVoice releases

2. Piper (ONNX models)

Speed: Fast | Quality: Good | Offline

Neural TTS engine powered by ONNX Runtime. Offers excellent voice quality with fast generation — processes text 10-50x faster than real-time on most CPUs.

Russian voices (4):

• Denis, Dmitri, Ruslan (Male)
• Irina (Female)

English voices (29):

US voices: Amy, Kathleen, Kristin, HFC Female, LJSpeech (Female) • Arctic, Bryce, Danny, HFC Male, Joe, John, Kusal, L2Arctic, Lessac, LibriTTS, Norman, Reza Ibrahim, Ryan, Sam (Male)

GB voices: Alba, Cori, Jenny Dioco, Southern English Female (Female) • Alan, Aru, Northern English Male, Semaine, VCTK (Male)

Download models: Piper Voices

3. Silero (PyTorch)

Speed: Medium | Quality: Excellent | Offline

Advanced neural TTS engine built on PyTorch. Delivers natural, expressive speech with excellent prosody.

Russian voices (5):

• Aidar, Eugene (Male)
• Baya, Kseniya, Xenia (Female)

English voices (4):

• Male 1, Male 2 (Male)
• Female 1, Female 2 (Female)

Models download automatically on first use (~100-200 MB).

4. Coqui XTTS-v2

Speed: Slow | Quality: Premium | Offline

State-of-the-art multilingual model with 14 built-in speaker voices. Produces the most natural-sounding speech among local engines with exceptional emotional range and prosody.

Voices (14, same for all languages):

Female: Claribel Dervla, Daisy Studious, Gracie Wise, Tammie Ema, Alison Dietlinde, Ana Florence, Annmarie Nele, Asya Anara

Male: Andrew Chipper, Badr Odhiambo, Dionisio Schuyler, Royston Min, Viktor Eka, Abrahan Mack

Supports 17 languages including Russian, English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Dutch, Czech, Arabic, Chinese, Japanese, Hungarian, Korean, and Hindi.

5. Bark Small

Speed: Slow | Quality: Experimental | Offline

Suno Bark Small is a generative text-to-audio model for expressive short-form speech. It supports Russian and English speaker presets and can produce pauses, laughter, sighs and other non-verbal sounds. Bark is best treated as experimental for audiobooks because it can deviate from the exact source text more than conventional TTS models.

Russian voices:

• 10 Bark Russian speaker presets

English voices:

• 10 Bark English speaker presets

Model downloads during Bark setup. GPU acceleration is supported through CUDA or DirectML when available, with CPU fallback for DirectML generation issues.

6. ElevenLabs (Cloud API)

Speed: Fast | Quality: Premium | Online

Premium cloud-based TTS with cutting-edge AI voice synthesis. Offers studio-quality output with remarkable naturalness.

Russian voices:

• Adam (Male)
• Rachel (Female)

English voices:

• Adam, Josh, Sam (Male)
• Rachel, Domi, Bella (Female)

Setup:

1. Create account at ElevenLabs
2. Get API key from your profile
3. Add to .env file:

ELEVENLABS_API_KEY=your_api_key_here

Installation

Prerequisites

• Node.js 18+
• Windows 10/11
• Python 3.9+ (for Silero and Coqui)

Quick Start

1. Clone repository

git clone <repo-url>
cd voicecraft

2. Install dependencies

npm install

3. Setup TTS components

Run the universal setup script:

# Via npm (recommended)
npm run setup

# Or directly via PowerShell
powershell .\scripts\setup-all.ps1

This will install:

• Piper TTS
• FFmpeg
• Silero TTS
• Coqui XTTS-v2
• Bark Small

4. Download voice models

For RHVoice:

Download and install from RHVoice releases. Voices will be automatically detected after installation.

For Piper:

Download voices from Piper releases and extract to:

tts_resources/piper/voices/

Structure:

tts_resources/
  piper/
    voices/
      ru_RU/
        denis/
          medium/
            ru_RU-denis-medium.onnx
            ru_RU-denis-medium.onnx.json
      en_US/
        lessac/
          medium/
            en_US-lessac-medium.onnx
            en_US-lessac-medium.onnx.json

For Silero:

Models download automatically on first use (~100-200 MB).

# Install only Silero
npm run setup:silero

For Coqui XTTS-v2:

Model downloads automatically on first use (~2 GB). Requires Python 3.9+ and GPU recommended for faster generation.

For Bark Small:

Model downloads during setup. Requires Python 3.9+ and GPU recommended for faster generation. Use shorter previews first because Bark is generative and may not follow long text exactly.

For ElevenLabs:

Add your API key to .env file:

ELEVENLABS_API_KEY=your_api_key_here

Usage

Development mode

npm run dev

Build application

npm run build
npm run package

Project Structure

voicecraft/
├── electron/               # Electron main process
│   ├── main.ts            # Main process entry point
│   ├── preload.ts         # IPC bridge (preload script)
│   ├── main/
│   │   ├── window.ts      # Window management
│   │   └── handlers/      # IPC handlers
│   └── services/
│       ├── parser.ts      # Book parsing
│       ├── setup/         # Dependency installation
│       └── tts/           # TTS services
├── src/                   # React frontend
│   ├── App.tsx           # Main component
│   ├── i18n/             # Internationalization (EN/RU)
│   ├── components/       # UI components
│   ├── hooks/            # React hooks
│   ├── fsm/              # State machine
│   └── utils/            # Utility functions
├── tts_resources/        # TTS resources
│   ├── piper/           # Piper TTS
│   ├── silero/          # Silero TTS
│   ├── coqui/           # Coqui XTTS-v2
│   ├── ffmpeg/          # FFmpeg for conversion
│   └── tts_server.py    # Python TTS server
├── scripts/
│   ├── setup-all.ps1    # Universal setup
│   ├── setup-silero.ps1 # Setup only Silero
│   └── release.cjs      # Release automation
└── .env                 # Environment variables (API keys)

Performance

GPU Acceleration

Silero and Coqui support hardware acceleration for faster speech generation:

CUDA is available for Silero, Coqui and Bark. DirectML is available for Coqui XTTS-v2 and Bark Small.

Enabling GPU Acceleration

1. During initial setup: When installing Silero, Coqui or Bark, select your preferred accelerator (CUDA, DirectML, or CPU). DirectML appears for Coqui and Bark when an AMD Radeon GPU is detected.

2. Change accelerator later: Go to Settings → TTS Setup → click "Reinstall" button next to Silero or Coqui to change the accelerator

Requirements

For NVIDIA CUDA:

• NVIDIA GPU with CUDA support (GTX 10xx or newer recommended)
• Latest NVIDIA drivers installed
• Automatically detected via nvidia-smi

For AMD DirectML:

• AMD Radeon GPU on Windows
• Latest AMD Adrenalin drivers installed
• DirectML dependencies are installed automatically
• Coqui voice cloning runs on CPU for cloned-voice requests when DirectML is selected, because XTTS voice cloning uses operations that DirectML does not support
• Bark DirectML support is experimental; if a DirectML generation operation fails, the Bark script retries on CPU

Auto-Detection

The application automatically detects available GPUs:

• Priority order: CUDA → DirectML → CPU
• GPU name and VRAM are displayed in the setup dialog
• If no compatible GPU is found, CPU mode is used

Parallelization

• RHVoice: up to 30 parallel threads
• Piper: up to 10 parallel threads
• Silero: up to 5 parallel threads
• Coqui: 1 thread (sequential processing)
• ElevenLabs: up to 3 parallel requests

Troubleshooting

RHVoice not working

• Install RHVoice from official releases
• Restart application after installation
• Voices are detected automatically via Windows SAPI

Piper not working

• Make sure voice models are downloaded
• Check directory structure
• .onnx and .onnx.json files must be in same folder

Silero slow generation

• Normal - it uses PyTorch models
• First run downloads models (~100-200 MB)
• For large books consider Piper or RHVoice

Coqui XTTS-v2 issues

• First run downloads ~2 GB model
• GPU recommended for faster generation
• Requires Python 3.9+
• Check that tts_resources/coqui/venv exists

Bark Small issues

• Bark is generative and may deviate from long source text
• Use short previews to evaluate a speaker preset before converting a book
• DirectML support is experimental; CPU fallback can be slower on AMD
• Check that tts_resources/bark-*/python/python.exe and generate.py exist

ElevenLabs not working

• Check that API key is set in .env file
• Verify API key is valid at elevenlabs.io
• Check internet connection

FFmpeg errors

• Make sure FFmpeg is installed: npm run setup
• Check that tts_resources/ffmpeg/ffmpeg.exe exists

License

MIT

Acknowledgements

• RHVoice - Lightweight SAPI voices
• Piper - Fast ONNX TTS models
• Silero - Natural PyTorch voices
• Coqui TTS - State-of-the-art XTTS-v2 model
• Suno Bark - Generative text-to-audio model
• ElevenLabs - Premium cloud TTS
• FFmpeg - Audio conversion