This repository contains a Marathi neural vocoder based on the HiFi-GAN architecture, trained for high-quality speech waveform generation from acoustic features.
๐ Overview
๐ Model Type: Neural Vocoder (HiFi-GAN)
๐ฃ๏ธ Language: Marathi
๐ฏ Task: Speech Resynthesis / Waveform Generation
๐ง Architecture: HiFi-GAN (GAN-based vocoder)
This model is trained to generate natural-sounding Marathi speech from mel-spectrogram inputs.
โ ๏ธ License
This model is released under:
Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
๐ Acknowledgements
This work is derived from:
The speech-resynthesis project by Meta (Facebook Research)
HiFi-GAN vocoder architecture
๐ Training Details
Model: HiFi-GAN
Training Type: Trained on Marathi speech data
Input: Mel-spectrograms
Output: Time-domain waveform
๐ฃ Citation
If you use this model, please cite:
@misc{marathi_hifigan_vocoder,
title={Marathi HiFi-GAN Vocoder},
author={ShrirajGaikwad},
year={2026},
note={CC BY-NC 4.0 Lโฆ