Divehi TTS – Female Voice (VITS-based)

This is a fine-tuned VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) model for Divehi speech synthesis. The model produces Female voice audio from Thaana-scripted Divehi text. Fine-tuned from Meta’s MMS-TTS architecture using a curated dataset of synthetic Divehi speech.

Model Details

Usage

from transformers import VitsModel, VitsTokenizer
import torchaudio

tokenizer = VitsTokenizer.from_pretrained("alakxender/mms-tts-div-ft-spk01-f01")
model = VitsModel.from_pretrained("alakxender/mms-tts-div-ft-spk01-f01")

text = "މޫސުން ވަރަށް ގޯސްވެ، ފުވައްމުލަކުން ފެށިގެން އައްޑުއަށް އޮރެންޖް އެލާޓް ނެރެފި"
inputs = tokenizer(text, return_tensors="pt")
waveform = model.generate(**inputs).waveform[0]

torchaudio.save("output.wav", waveform.unsqueeze(0), 16000)

Evaluation Summary

• Model: alakxender/mms-tts-div-ft-spk01-f01
• Evaluated Samples: 3
• Avg Estimated MOS (UTMOS): 3.198

  {
    "5": "Excellent (very natural)",
    "4": "Good (mostly natural)",
    "3": "Fair (some robotic quality)",
    "2": "Poor (noticeably unnatural)",
    "1": "Bad (unintelligible or very synthetic)"
  }
  

• Artifacts:
  • 🎵 Audio: outputs/audio/
  • 📊 Spectrograms: outputs/spectrograms/
  • 📄 Report: outputs/report.txt
  • 📈 MOS Scores: outputs/mos_scores.txt

Acknowledgements

• Meta MMS-TTS
• Tarepan's SpeechMOS
• Hugging Face 🤗 Transformers