Survey On Monolingual Speech-to-Speech Translation
Yash Vachhani, Prashant J. Viradiya · International Journal of Research in Advent Technology · 2025
Direct Speech-to-Speech (S2S) translation represents a significant goal in facilitating seamless cross- lingual communication, aiming to overcome the latency and error propagation issues inherent in traditional cascaded systems (ASRMT-TTS). However, the development of high- performing direct S2S models has been critically constrained by the scarcity of large-scale parallel S2S corpora. This survey details Translatotron 3, a pivotal direct S2S system introduced by Google Research in 2023. Translatotron 3 fundamentally shifts the paradigm by demonstrating, for the first time, the feasibility of training a high-quality, end-toend S2S model exclusively using readily available monolingual data resources: source/target speech and source/target text. Leveraging innovative techniques such as unsupervised utterance splitting, phoneme-based intermediate representations, speech-adapted back-translation, and a non auto regressive decoder for rapid inference, Translatotron 3 achieves strong translation quality and remarkable speaker voice preservation without requiring any parallel S2S examples. We critically review the technological context, dissect the model’s architecture and training methodology, analyze its reported performance benchmarks, and discuss its profound implications for advancing S2S translation, particularly for the vast number of low-resource languages previously underserved by data-hungry models.