SPEECH TRANSLATION IN HOST VOICE, TONE

Abstract

This research paper presents a novel Powerful, Efficient , Reliable approach to speech translation that preserves the host's original voice characteristics including tone, pitch, emotional cues, and speaking style. We propose a multi-head attention mechanism specifically designed for phonetic features that operates alongside traditional speech recognition and translation pipelines. Our method extracts 14 or more distinct speech characteristics and applies specialized attention heads to each, enabling the translated speech to maintain the original speaker's vocal identity while converting language content.