Smol Urdu TTS V1

Urdu Text-to-Speech model based on SmolLM2-135M.

Note: This is an experimental model and meant as a starting point for the Urdu community. Quality may not be perfect, and improvements are coming soon.

Audio Samples

Below are sample audio outputs generated by Smol Urdu TTS V1 for different Urdu text inputs.

Training Metrics Snapshot

Figure: Training dynamics illustrating loss convergence, learning rate decay, and stable gradient norms during fine-tuning.

Usage

https://github.com/Proxima-AI-Co/smolurdu

Model Details

• Base Model: SmolLM2-135M-Instruct
• Parameters: 135M
• Language: Urdu
• Sample Rate: 24kHz
• Checkpoint: 120000 steps

Citation

@misc{smol-urdu-tts-v1,
  title={Smol Urdu TTS V1: Urdu Text-to-Speech},
  author={Adnan Zaidi, Mahwiz Khalil},
  year={2024},
  howpublished={HuggingFace Model Repository}
}