NOTE: since the original repo at https://github.com/alasdairforsythe/kokoro-voice-composer was taken down and it used MIT license, I guess it's okay to upload it here for people to use.

Kokoro Voice Composer

Kokoro Voice Composer is a Python Streamlit UI for creating and customizing text-to-speech (TTS) voices for the Kokoro TTS model. Blend existing voices, apply transformations, and export them for use with kokoro and kokoro-onnx. It will run on CPU or GPU.

View the example on Hugging Face Spaces.

---

Installation

1. Clone this repository:

   git clone https://github.com/alasdairforsythe/kokoro-voice-composer.git
   cd kokoro-voice-composer

2. Install the required Python packages using pip:

   pip install kokoro-onnx onnxruntime misaki[en,zh,ja] streamlit pydub soundfile torch numpy ordered_set pypinyin cn2an jieba fugashi jaconv mojimoji

   For GPU add kokoro-onnx[gpu]

3. Ensure eSpeak is installed (e.g., on Ubuntu: sudo apt install espeak-ng). For GPU also install portaudio19-dev

4. Run the application:

   streamlit run voice-composer.py --server.port 5544 --server.fileWatcherType none

   The app will automatically download available voice models upon first startup. Refresh the page to get rid of the notifications.

---

Features

• Voice Blending: Combine multiple voice embeddings with adjustable weights.
• Seed Transformation: Add controlled noise to create unique voice variations.
• Inflection Transfer: Transfer vocal inflections from one voice to another.
• PCA Perturbation and Spectral Transform: Other transformations.
• Export Options: Save voices as .pt files or .bin files compatible with kokoro-onnx.
• Real-Time Playback: Preview and download generated audio with customizable text, language, and speed.

---

Recommended Workflow

To efficiently create high-quality custom voices, follow this recommended workflow:

1. Voice Discovery:

   • Select an existing voice.
   • Experiment with different seeds using Seed Transform until a desired voice quality or inflection is achieved.
   • Save desirable voices; they will appear in your 'Generated' list.

2. Voice Refinement:

   • If the newly created voice has distortions, blend it lightly (weights around 0.1 - 0.2) with similar existing voices to smooth out imperfections.

3. Inflection Optimization:

   • Replace the voice’s inflection with a more preferred option using Inflection Transfer. Keep the default strength of 0.3 for optimal results.

4. Reusing Inflections:

   • Store any seed transformations with good inflections for future use in other voices.

---

Using Exported Voices with kokoro-onnx

Exported ONNX voices can be integrated easily:

kokoro = Kokoro("kokoro-v1.0.onnx", "your_voice.bin")
samples, sample_rate = kokoro.create("Yes! It works!", voice="voice", speed=1.0, lang="en-us")

License

This project is licensed under the MIT License.