Multilingual and Accent-Aware Transformation of Read Speech to Conversational Speech

This paper introduces a framework that brings together statistical modeling and speech synthesis methods to convert read speech into spoken conversational speech. To generate expressive, natural-sounding speech, the system employs controlled pauses, prosody modulation, and filler word injection. To simulate human-like conversational tendencies, the model alters pitch, tempo, and intonation using the Google Cloud Text-to-Speech API with SSML-based changes. With the help of tools such as Praat for acoustic parameter extraction, there were over 300 speech samples that were examined during the evaluation. The mean filler pause duration was 0.27 seconds (SD: 0.26 seconds), while the mean duration for non-filler pauses was 0.09 seconds (SD: 0.18 seconds). The Mean Opinion Score (MOS) of the system was in between 4.0 to 4.2, reflecting enhanced naturalness of speech.Its performance, as compared to traditional text-to-speech systems, was also confirmed by evaluating performance via waveform analysis and Jaccard similarity. These results show the potential uses of the system in assistive technology, content reading, and virtual assistants. For increased flexibility, upcoming research will also explore real-time deployment on edge devices and combination with neural text-to-speech models.