Swa-CSM-1B

Swa-CSM-1B is the best open-source Swahili text-to-speech model, fine-tuned from Sesame's CSM-1B for high-quality Swahili speech generation. This model powers Soga, a Swahili AI voice assistant.

Developed by Nadhari AI Lab .

Model Details

• Base Model: sesame/csm-1b
• Language: Swahili (sw)
• Task: Text-to-Speech
• Architecture: Llama backbone with Mimi audio decoder (RVQ codes)

Usage

Basic Generation

import torch
from transformers import CsmForConditionalGeneration, AutoProcessor

model_id = "Nadhari/swa-csm-1b"
device = "cuda" if torch.cuda.is_available() else "cpu"

processor = AutoProcessor.from_pretrained(model_id)
model = CsmForConditionalGeneration.from_pretrained(model_id, device_map=device)

# Generate Swahili speech
text = "[0]Mambo vipi?."
inputs = processor(text, add_special_tokens=True).to(device)

audio = model.generate(**inputs, output_audio=True)
processor.save_audio(audio, "swahili_output.wav")

With Conversation Context

For best results, provide conversational context:

conversation = [
    {"role": "0", "content": [{"type": "text", "text": "Habari yako?"}]},
]
inputs = processor.apply_chat_template(
    conversation,
    tokenize=True,
    return_dict=True,
).to(device)

audio = model.generate(**inputs, output_audio=True)
processor.save_audio(audio, "swahili_conversation.wav")

Intended Use

This model is designed for:

• Swahili text-to-speech applications
• Research in African language speech synthesis.

Limitations

• Optimized for standard Swahili; performance may vary with regional dialects
• Inherited limitations from base CSM-1B model
• Best results achieved with conversational context

Ethical Considerations

This model should not be used for:

• Impersonation or fraud
• Generating misleading or deceptive content
• Any illegal or harmful activities

By using this model, you agree to use it responsibly and ethically.

Citation

@misc{swa-csm-1b,
  author = {Nadhari AI Lab},
  title = {Swa-CSM-1B: Swahili Text-to-Speech Model},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/Nadhari/swa-csm-1b}
}

Acknowledgments

• Sesame for the base CSM-1B model

Contact

• Organization: Nadhari AI Lab
• Author: Alfaxad
• Email: [research@nadhari.ai]

---