YarnGPT Python Wrapper Library

Version 0.2.0

• Added support for local speakers
• Added support for multiple languages
• Added support for multiple languages

Description

YarnGPT is a Python wrapper for the YarnGPT text-to-speech model, designed to synthesize natural Nigerian speech in multiple languages using a pure language modeling approach. This library provides a simple API to convert text into audio output, allowing users to select from various preset voices, languages, and adjust generation parameters.

Features

• Supports multiple Nigerian languages: English, Yoruba, Igbo, and Hausa
• Rich set of voices for each language:
  • English: idera, chinenye, jude, emma, umar, joke, zainab, osagie, remi, tayo
  • Yoruba: abayomi, aisha, folake
  • Igbo: chioma, obinna, adanna
  • Hausa: amina, fatima, ibrahim, yusuf
• Utilizes Hugging Face's model caching for efficient model loading
• Exposes a straightforward API function: generate_speech(text, speaker, language, temperature, repetition_penalty, max_length)
• Allows customization of generation parameters
• Includes unit tests to ensure core functionality

Installation

1. Create and activate a virtual environment:

   • On Linux/MacOS:

   python3 -m venv env
   source env/bin/activate
   

   • On Windows:

   python -m venv env
   env\Scripts\activate
   

2. Install the package:

   pip install yarngpt
   

Usage

Basic usage to generate and save audio:

from yarngpt import generate_speech
import torchaudio

# Generate English speech with default speaker
audio = generate_speech("Hello, this is a test.", language="english")

# Generate Yoruba speech with a Yoruba voice
audio = generate_speech("Bawo ni?", speaker="abayomi", language="yoruba")

# Save the generated audio
torchaudio.save("output.wav", audio, sample_rate=24000)

For Jupyter Notebook users, you can also play the audio directly:

from yarngpt import generate_speech
import torchaudio
from IPython.display import Audio

# Generate speech in different languages
english_audio = generate_speech("Hello!", speaker="idera", language="english")
yoruba_audio = generate_speech("Bawo ni?", speaker="abayomi", language="yoruba")
igbo_audio = generate_speech("Kedu?", speaker="chioma", language="igbo")
hausa_audio = generate_speech("Sannu!", speaker="amina", language="hausa")

# Save and play the audio
torchaudio.save("output.wav", english_audio, sample_rate=24000)
Audio("output.wav")

Parameter Options

• text: The input string to convert to speech
• speaker: Choose from available voices by language (see Features section for full list)
• language: The language for speech generation ("english", "yoruba", "igbo", "hausa")
• temperature: Controls the randomness of generation (default is 0.1)
• repetition_penalty: A factor to reduce repetitive output (default is 1.1)
• max_length: The maximum length of the generated output tokens (default is 4000)

Testing

Run the unit tests to verify functionality:

python -m unittest discover -s tests

License

This project is licensed under the MIT License.

Acknowledgments

• Built as a contribution to yarngpt projects
• Utilizes Hugging Face's model caching and the transformers library
• Special thanks to the open-source community for their ongoing support

For more details and documentation, visit the GitHub repository: https://github.com/jerryola1