Multi-Step TTS Architecture for Natural and Comprehensible Speech Synthesis from Improved Summaries

Abstract

Text-to-Speech (TTS) technology has a central role in enabling accessibility of content for visually impaired students. Traditional TTS engines, however, are likely to fail in clarity, prosody, and coherence when translating machine-generating summaries into speech, especially if the text incorporates intricate syntactic forms or mathematical equations. The paper introduces a multi-stage model designed to encourage naturalness and intelligibility for educational content narration. The model integrates DistilBERT-based linguistic embeddings for contextual text comprehension, FastSpeech 2 for real acoustics modeling and prosody adaptation, and HiFi-GAN for high-fidelity waveform generation. Through sentence continuity and prosodic transition refinement, the model offers smoother and more coherent speech generation for learning. Experimental testing verifies that the hybrid model proposed performs better than baseline models. In contrast to Tacotron with HiFi-GAN (MOS = 3.2, WER = 12.8%) and FastSpeech 2 with WaveGlow (MOS = 3.8, WER = 9.5%), the DistilBERT + FastSpeech 2 + HiFi-GAN model has a Mean Opinion Score (MOS) of 4.4 and a Word Error Rate (WER) of 6.3%, showing a drastic improvement in both perceptual quality and linguistic accuracy. The findings verify that the use of transformer-based semantic modeling and deep vocoding greatly improves speech clarity, fluency, and naturalness.