9jaLingo/9jaLingo-TTS-African-ckpt-25k

Spark-TTS African Voice

Text-to-speech model for African languages, fine-tuned using the Spark-TTS / SparkVox framework.

Languages Supported

• Igbo (ig)
• Yoruba (yo)
• Hausa (ha)
• Nigerian Pidgin (pcm)

Features

• Text-to-Speech (TTS): Generate natural speech from text
• Voice Cloning: Clone any voice from a short reference audio
• Controllable TTS: Control pitch, speed, and gender

Model Structure

This model follows the official Spark-TTS HuggingFace format:

├── config.yaml
├── BiCodec/
│   ├── config.yaml
│   └── model.safetensors
├── LLM/
│   ├── config.json
│   ├── model.safetensors
│   └── (tokenizer files)
└── wav2vec2-large-xlsr-53/
    ├── config.json
    ├── preprocessor_config.json
    └── pytorch_model.bin

Usage with Spark-TTS Inference

# 1. Clone the official Spark-TTS inference repo
# git clone https://github.com/SparkAudio/Spark-TTS.git
# cd Spark-TTS

# 2. Download this model
# from huggingface_hub import snapshot_download
# snapshot_download("NaijaLingo/Spark-TTS-African", local_dir="pretrained_models/Spark-TTS-African")

# 3. Run inference
from cli.SparkTTS import SparkTTS

model = SparkTTS(model_dir="pretrained_models/Spark-TTS-African", device="cuda:0")

# Voice cloning
wav = model.inference(
    text="How you dey, hope say you dey fine?",
    prompt_speech_path="path/to/reference.wav",
    prompt_text="Optional transcript of reference audio",
)

# Controllable TTS (no reference audio needed)
wav = model.inference(
    text="Sannu da zuwa, yaya kake?",
    gender="female",
    pitch="moderate",
    speed="moderate",
)

Training Details

Component	Details
Base LLM	Qwen2.5-0.5B-Instruct
Audio Tokenizer	BiCodec (fine-tuned on African voice data)
Feature Extractor	wav2vec2-large-xlsr-53
Dataset	chukypedro/african_voice
Training Framework	SparkVox

License

Apache 2.0