---
license: apache-2.0
language:
- mos
- fr
tags:
- text-to-speech
- tts
- moore
- burkina-faso
- african-languages
- voice-cloning
- speech-synthesis
pipeline_tag: text-to-speech
---

# BIA-SPARKTTS-V2

## Model Description

BIA-SPARKTTS-V2 is a fine-tuned version of SparkTTS specifically optimized for Moore language text-to-speech synthesis. This model enables high-quality speech generation in Moore with voice cloning capabilities, supporting the linguistic needs of Burkina Faso and the wider Moore-speaking community.

Our model is built upon the powerful SparkTTS architecture, leveraging state-of-the-art neural vocoding and voice cloning techniques. It has been fine-tuned on curated Moore speech data to ensure natural and contextually appropriate speech synthesis for Moore speakers.

## Language Information

- Primary Language: Moore (Mooré)
- Language Code: `moor_Latn`
- Speech Synthesis: Text-to-Speech (TTS)
- Region: Burkina Faso, West Africa
- Speakers: ~5 million native speakers
- Language Family: Niger-Congo → Gur → Oti-Volta → Moore

## Model Performance

The model achieves the following results on the evaluation set:

- Step 1000: Training Loss: 4.158300 | Validation Loss: 4.068496
- Step 2000: Training Loss: 3.816300 | Validation Loss: 3.789840
- Step 3000: Training Loss: 3.655900 | Validation Loss: 3.688933
- Step 4000: Training Loss: 3.609400 | Validation Loss: 3.628450
- Step 5000: Training Loss: 3.528000 | Validation Loss: 3.593397
- Step 6000: Training Loss: 3.463600 | Validation Loss: 3.563343
- Step 7000: Training Loss: 3.404400 | Validation Loss: 3.546153
- Step 8000: Training Loss: 3.395500 | Validation Loss: 3.528718
- Step 9000: Training Loss: 3.367900 | Validation Loss: 3.524420
- Step 10000: Training Loss: 3.359900 | Validation Loss: 3.522905
- Parameters: 0.5B (500M)
- Total Training Steps: 10,680 (8 epochs)
- Sample Rate: 16kHz
- Audio Quality: Natural, clear pronunciation

## Features

- **Voice Creation**: Generate speech with customizable gender, pitch, and speed
- **Voice Cloning**: Clone voices from audio samples (male/female)
- **Custom Voice Cloning**: Upload your own reference audio for personalized voice cloning
- **Flexible Generation**: Adjustable temperature, top-k, and top-p sampling parameters

## Usage

### Installation

```bash
pip install torch transformers gradio soundfile
pip install git+https://github.com/SparkAudio/SparkTTS.git
```

### Basic Usage

```python
import torch
from SparkTTS import SparkTTS

# Initialize model
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model = SparkTTS("burkimbia/BIA-SPARKTTS-V1", device=device)

# Generate speech with voice creation
text = "Yʋʋmd fãa, a ɛlkana rag n yita a tẽngẽ wã"
wav = model.inference(
    text,
    gender="male",
    pitch="moderate",
    speed="moderate",
    do_sample=True,
    temperature=0.8,
    top_k=50,
    top_p=0.95
)

# Save audio
import soundfile as sf
sf.write("output.wav", wav, 16000)
```

### Voice Cloning with Audio Sample

```python
# Clone voice from reference audio
reference_audio = "path/to/reference/audio.wav"
wav = model.inference(
    text,
    prompt_speech_path=reference_audio,
    prompt_text=None,  # Optional: transcript of reference audio
    do_sample=True,
    temperature=0.8,
    top_k=50,
    top_p=0.95
)
```

### Advanced Parameters

```python
# Fine-tune generation with custom parameters
wav = model.inference(
    text,
    gender="female",
    pitch="high",           # very_low | low | moderate | high | very_high
    speed="moderate",       # very_low | low | moderate | high | very_high
    do_sample=True,         # Enable sampling for more natural variation
    temperature=0.8,        # Higher = more creative (0.1-1.5)
    top_k=50,              # Number of tokens to consider (1-100)
    top_p=0.95,            # Nucleus sampling threshold (0.1-1.0)
)
```

## Model Details

### Training Configuration
- Base Model: SparkTTS-0.5B
- Model Size: 500M parameters
- Training Steps: 10,680 steps (8 epochs)
- Batch Size: 32 per device
- Gradient Accumulation Steps: 2
- Effective Batch Size: 64
- Learning Rate: 5e-5
- LR Scheduler: Cosine with 500 warmup steps
- Optimizer: AdamW 8-bit with 0.01 weight decay
- Sample Rate: 16kHz
- Max Tokens: 3000
- Hardware: NVIDIA A100-SXM4-40GB (40GB VRAM)
- Framework: PyTorch (FP32)
- Seed: 2025

### Training Progress
The model showed consistent improvement during training:
- Rapid initial improvement (4.07 → 3.79 validation loss in first 2000 steps)
- Steady convergence from 3.69 to 3.52 over 10,000 steps
- Best model checkpoint saved at step 10,000 (validation loss: 3.522905)
- Evaluation frequency: Every 1000 steps
- Total training time: ~12-14 hours on A100

### Architecture Components
- **Audio Tokenizer**: BiCodec for high-quality audio encoding
- **Language Model**: Causal LM for sequential token generation
- **Vocoder**: Neural vocoder for waveform synthesis
- **Voice Encoder**: For voice cloning and style transfer

## Generation Parameters

### do_sample
- `True`: Enables sampling for natural variation (recommended)
- `False`: Greedy decoding for deterministic output

### temperature
- Range: 0.1 - 1.5
- Default: 0.8
- Lower values (0.3-0.5): More deterministic, consistent pronunciation
- Higher values (0.8-1.2): More natural variation and expressiveness

### top_k
- Range: 1 - 100
- Default: 50
- Number of highest probability tokens to consider during sampling

### top_p (nucleus sampling)
- Range: 0.1 - 1.0
- Default: 0.95
- Cumulative probability threshold for token selection

## Limitations

- The model performs best on standard Moore text with proper orthography
- Extremely long texts may require chunking
- Voice quality depends on reference audio quality for cloning
- Generated speech may vary slightly between runs when sampling is enabled
- Specialized technical terms may require phonetic adjustments

## Use Cases

- **Education**: Moore language learning and pronunciation
- **Accessibility**: Text-to-speech for visually impaired Moore speakers
- **Content Creation**: Audiobooks, podcasts, and media in Moore
- **Communication**: Voice assistants and chatbots for Moore speakers
- **Preservation**: Digital archiving of Moore language resources

## Gradio Demo

A web interface is available for easy testing:

```python
import gradio as gr
from SparkTTS import SparkTTS

model = SparkTTS("burkimbia/BIA-SPARKTTS-V1")

def generate_speech(text, gender, pitch, speed):
    wav = model.inference(text, gender=gender, pitch=pitch, speed=speed)
    return (16000, wav)

demo = gr.Interface(
    fn=generate_speech,
    inputs=[
        gr.Textbox(label="Moore Text"),
        gr.Radio(["male", "female"], label="Gender"),
        gr.Slider(1, 5, step=1, value=3, label="Pitch"),
        gr.Slider(1, 5, step=1, value=3, label="Speed"),
    ],
    outputs=gr.Audio(label="Generated Speech")
)

demo.launch()
```

## Feedback

We're continuously striving to improve our model's performance and usability. If you have any feedback, suggestions, or encounter any issues, please don't hesitate to reach out to us at [BurkimbIA](https://huggingface.co/burkimbia).

## Citation

If you use this model in your research, please cite:

```bibtex
@misc{bia-sparktts-v1,
  title={BIA-SPARKTTS-V2: A Fine-tuned SparkTTS Model for Moore Language Speech Synthesis},
  author={BurkimbIA},
  year={2025},
  howpublished={\url{https://huggingface.co/burkimbia/BIA-SPARKTTS-V1}}
}
```

## Acknowledgments

- Coqui tts team for the SparkTTS architecture and pretrained models
- Speech data contributors who provided high-quality recordings
- BurkimbIA team for data curation and model development

## Framework Versions

- Transformers: 4.57.1
- PyTorch: 2.2.1+cu121
- Python: 3.12+
- Gradio: 5.0+


Developed by [BurkimbIA](https://huggingface.co/burkimbia) - Advancing African Language Technologies