Swar (Hindi): A Speaker-Aware Vocoder for Low-Resource Speech Synthesis

Hindi, one of the most widely spoken languages, is considered low-resource for speech synthesis (SS) due to its complex phonetic structure and limited annotated datasets. In this paper, we propose Swar (Hindi), a speaker-aware neural vocoder designed specifically for Hindi. Swar (Hindi) uses a prosodic adaptation mechanism and speaker-aware conditioning to address challenges, such as poor generalisation and context-sensitive pronunciation in prosody. Swar (Hindi) is capable of generating high-quality 22.05 kHz speech with a GPU inference speed of 6.53× real-time factor (RTF) on an NVIDIA GTX 1080 8GB VRAM by offering a practical balance between inference quality and performance. Generated samples evaluation using both subjective (MOS, SMOS) and objective metrics, such as PCC, PESQ, STOI, MSD, and MCD to evaluate Swar (Hindi)s' effectiveness over baseline GAN-based vocoders. Swar (Hindi) achieved a MOS score improvement of +0.34 over the baseline by enhancing overall speech quality. Additionally, a +0.14 gain in SMOS score by showcasing better preservation of prosody and naturalness in speech synthesis, even under low-resource constraints. Swar (Hindi) achieved MOS score +0.34 higher than baseline, and the SMOS score of +0.14 showed preservation of prosody and naturalness in speech synthesis under low-resource constraints.