Vaachika: A GAN-Based Neural Vocoder for Marathi Language

Vaachika Marathi introduces a neural vocoder for high-fidelity speech synthesis. Our proposed model presents a significant enhancement to a baseline GAN-based vocoder by focusing on improving speech quality and training stability. The proposed model includes 78 layers within the generators' architecture in order to capture complex linguistic features and preserve phonetic features (e.g., intonations, vowel nasalizations, and pronunciations). Additionally, we adopted the ReLU6 activation function to reduce the risk of exploding gradients and to enhance training convergence and stability. Generated samples at a 22.04 kHz sampling rate and evaluated them with subjective, objective, and quantitative measures to assess their fidelity and perceptual quality. The generated samples from the Vaachika V1 model achieved a 4.65 on the 5-scale Mean Opinion Score (MOS), with a significant improvement over the baseline model's score of +0.29. The proposed Vaachika V2 model achieves a 51.10% reduction in training time on an NVIDIA Quadro XP100 16 GB VRAM GPU, while improving audio quality metrics by up to +2.006 (in terms of PESQ), and reducing spectral distortion (MCD) by 75.6%, compared to the baseline.