README.md

metadata

library_name: transformers
tags:
  - unsloth
  - trl
  - sft
  - tts
license: apache-2.0
language:
  - ary
datasets:
  - KandirResearch/DarijaTTS-clean
base_model:
  - OuteAI/OuteTTS-0.2-500M
pipeline_tag: text-to-speech

Moroccan Darija TTS

This is a text-to-speech (TTS) model for Moroccan Darija, fine-tuned from OuteAI/OuteTTS-0.2-500M on the KandirResearch/DarijaTTS-clean dataset.

Model Details

• Base Model: OuteAI/OuteTTS-0.2-500M
• Dataset: KandirResearch/DarijaTTS-clean
• Training Method: Fine-tuned using Unsloth's SFTTrainer
• Dataset Preparation: Preprocessed following OuteTTS training guide
• Demo: Try it here

Usage

You can run the model using outetts as follows:

install outetts and llama-cpp-python

pip install outetts llama-cpp-python huggingface_hub

import outetts
from outetts.models.config import GenerationConfig
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="KandirResearch/DarijaTTS-v0.1-500M",
    filename="unsloth.Q8_0.gguf",
)
model_config = outetts.GGUFModelConfig_v2(
    model_path=model_path,
    tokenizer_path="KandirResearch/DarijaTTS-v0.1-500M",
)
interface = outetts.InterfaceGGUF(model_version="0.3", cfg=model_config)

def tts(text, temperature=0.3, repetition_penalty=1.1):
    gen_cfg = GenerationConfig(
        text=text,
        temperature=temperature,
        repetition_penalty=repetition_penalty,
        max_length=4096,
    )
    output = interface.generate(config=gen_cfg)
    output_path = "output.wav"
    output.save(output_path)
    return output_path

# Example usage
audio_path = tts("السلام كيداير لاباس عليك؟")
print(f"Generated audio saved at: {audio_path}")

Training

The model was fine-tuned using Unsloth's SFTTrainer. The dataset was preprocessed following the OuteTTS training guide. LoRA-based fine-tuning was applied to improve efficiency.

Support Me

---

For any issues or improvements, feel free to open a discussion or PR!