CrossVoice: Crosslingual Prosody Preserving Cascade-S2ST using Transfer Learning

TL;DR

CrossVoice tackles cross-lingual speech-to-speech translation with preserved prosody by adopting a cascade S2ST pipeline that integrates Faster-Whisper ASR, Google NMT, and MMS-based VITS-TTS, enhanced with a transfer-learning voice-cloning module using X-vectors for prosody transfer. The approach demonstrates competitive translation quality and naturalness compared to direct S2ST baselines, achieving a mean BLEU around the mid-30s to mid-40s across tasks and MOS scores approaching ground-truth speech. Experiments on Fisher Es-En, VoxPopuli Fr-En, CVSS-T, and IndicTTS show CrossVoice outperforms state-of-the-art direct S2ST in BLEU (notably a ~19-point gain on VoxPopuli Fr-En) and preserves cross-lingual prosody more effectively through the transfer-learning pathway. The work underscores the viability of cascade architectures with targeted prosody transfer as a practical alternative to end-to-end direct S2ST for multilingual speech interfaces, while acknowledging data requirements and ongoing challenges in prosody transfer across languages.

Abstract

This paper presents CrossVoice, a novel cascade-based Speech-to-Speech Translation (S2ST) system employing advanced ASR, MT, and TTS technologies with cross-lingual prosody preservation through transfer learning. We conducted comprehensive experiments comparing CrossVoice with direct-S2ST systems, showing improved BLEU scores on tasks such as Fisher Es-En, VoxPopuli Fr-En and prosody preservation on benchmark datasets CVSS-T and IndicTTS. With an average mean opinion score of 3.75 out of 4, speech synthesized by CrossVoice closely rivals human speech on the benchmark, highlighting the efficacy of cascade-based systems and transfer learning in multilingual S2ST with prosody transfer.

Figures (3)

• Figure 1: Proposed Architecture for CrossVoice
• Figure 2: BLEU scores on CVSS-T for X-en Translation
• Figure 3: BLEU scores on CVSS-T for en-X Translation