A Study on Intelligibility, Naturalness and Fluency of Text-to-Speech Systems

Text-to-Speech (TTS) technology advancements have resulted in the developing of sophisticated models like Tacotron and WaveNet. This research evaluates these models based on clarity, naturalness, and flow. Tacotron employs a comprehensive architecture that converts text into mel-spectrograms, excelling in prosody and fluency. WaveNet, a probabilistic model, directly generates raw audio waveforms, delivering exceptional sound quality. The LJSpeech dataset assessed performance through metrics such as Word Error Rate (WER) and Mean Opinion Score (MOS). Tacotron achieved a WER of 6.2%, indicating strong fluency with some alignment challenges. WaveNet outperformed Tacotron with a WER of 4.7% and a MOS of 4.6, demonstrating greater naturalness but some minor fluency issues. Tacotron is ideal for expressive applications, while WaveNet excels in realism. Future research could explore hybrid models to combine these strengths and address current limitations..