Whether to Use Pretrain Neural Vocoders or Not?: An Empirical Analysis for Indic Speech Synthesis

Existing speech synthesis systems are developed in German, Mandarin, or Japanese. However, languages, such as Indic and African lack such progress in building robust speech synthesis systems due to limited low-resource constraints (compared to language diversity). Manually creating a dataset specifically for these languages is time-consuming and costly. This paper conducts comprehensive experiments and evaluations of eight non-Indic pre-trained neural vocoders (PTNV) in six Indic languages and four dialects. We used subjective (NISQA Mean Opinion Score (MOS)), objective (Perceptual Evaluation of Speech Quality (PESQ), Mel-Cepstral Distortion (MCD), Mel-Spectra Distance (MSD), Short-Time Objective Intelligibility (STOI), Pearson Correlation Coefficient (PCC)) and quantitative (Word Error Rate (WER), Character Error Rate (CER)) measures to assess the generalization capabilities and performance of PTV models in Indic languages. During analysis, we noticed that architectural features, such as Multi-Receptive Fields (MRFs) and residual blocks, significantly improved the quality and intelligibility of the synthesized speech across diverse Indic languages. In particular, vocoders HiFi-GAN and UnivNet gave superior generalization and perceptual quality, despite being trained on non-indic datasets. Which suggests strong cross-lingual capabilities when combined with proper conditioning and architectural robustness.