Pocket TTS

Lightweight CPU-based Text-to-Speech extension for text-generation-webui using Kyutai's Pocket TTS model.

Features

• Lightweight: 100M parameter model (vs 1.6B for alternatives like Moshi)
• CPU-only: No GPU contention with your LLM
• Fast: 3-4x faster than realtime speech generation
• Simple: ~220 lines of code, no external servers
• 8 Built-in Voices: alba, marius, javert, jean, fantine, cosette, eponine, azelma
• Voice Cloning: Support for custom voice WAV files

Requirements

• text-generation-webui
• Python 3.10+
• numpy < 2.0 (required for text-generation-webui compatibility)

Installation

Option 1: Git Clone (Recommended)

cd /path/to/text-generation-webui/extensions
git clone https://github.com/jschuller/pocket_tts.git
pip install -r pocket_tts/requirements.txt

Option 2: Manual Download

1. Download and extract to text-generation-webui/extensions/pocket_tts/

2. Install dependencies:

   cd /path/to/text-generation-webui
   pip install -r extensions/pocket_tts/requirements.txt

3. Enable the extension by adding to user_data/CMD_FLAGS.txt:

   --extensions pocket_tts
   

4. Restart text-generation-webui

Configuration

The extension adds a "Pocket TTS" accordion in the Chat tab with these options:

Setting	Description	Default
Activate TTS	Enable/disable audio generation	On
Play automatically	Autoplay audio for new responses	On
Show message text	Display text below audio player	Off
Voice	Select from 8 built-in voices	alba

Custom Voices

To use your own voice for cloning:

1. Place a WAV file (10-30 seconds of clear speech) in the voices/ directory
2. The voice will appear in the dropdown menu
3. Select it and generate a response

Tips for good voice cloning:

• Use high-quality audio (16kHz+ sample rate)
• Clear speech without background noise
• 10-30 seconds of varied speech

Usage

1. Start text-generation-webui with the extension enabled
2. Load any LLM model
3. Start a chat - responses will automatically generate audio
4. Use the Preview button to test voice settings

API

The extension hooks into these text-generation-webui extension points:

• output_modifier() - Generates audio for bot responses
• state_modifier() - Disables streaming (required for TTS)
• history_modifier() - Removes autoplay from old messages
• ui() - Adds configuration UI

Troubleshooting

"numpy.dtype size changed" error

Install numpy < 2.0:

pip install "numpy<2.0"

First response is slow

The TTS model loads on first use (~2-3 seconds). Subsequent responses are fast.

No audio plays

• Check browser autoplay settings
• Verify the extension is enabled in Session tab
• Check console for errors

Audio quality issues

• Try different voices (some work better for certain content)
• Shorter responses generally sound better
• The model works best with English text

Technical Details

• Model: Kyutai Pocket TTS (100M parameters)
• Sample Rate: 24kHz
• Output Format: WAV
• Voice State Caching: Voice embeddings are cached for performance

Files

pocket_tts/
├── script.py           # Main extension code
├── requirements.txt    # Python dependencies
├── README.md           # This file
├── LICENSE             # MIT License
├── test_tts.py         # Standalone test script
├── test_sentences.txt  # Preview sentences
├── outputs/            # Generated audio files
└── voices/             # Custom voice WAV files

Credits

• Kyutai for Pocket TTS
• oobabooga for text-generation-webui
• Based on the silero_tts extension pattern

License

MIT License - see LICENSE file