MoVE: Translating Laughter and Tears via Mixture of Vocalization Experts in Speech-to-Speech Translation
Abstract
Recent Speech-to-Speech Translation (S2ST) systems achieve strong semantic accuracy yet consistently strip away non-verbal vocalizations (NVs), such as laughter and crying that convey pragmatic intent, which severely limits real-world utility. We address this via three contributions. First, we propose a synthesis pipeline for building scalable expressive datasets to overcome the data scarcity limitation. Second, we propose MoVE, a Mixture-of-LoRA-Experts architecture with expressive-specialized adapters and a soft-weighting router that blends experts for capturing hybrid expressive states. Third, we show pretrained AudioLLMs enable striking data efficiency: 30 minutes of curated data is enough for strong performance. On English-Chinese S2ST, while comparing with strong baselines, MoVE reproduces target NVs in 76% of cases and achieves the highest human-rated naturalness and emotional fidelity among all compared systems, where existing S2ST systems preserve at most 14% of NVs.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
This paper presents a pioneering approach to enhancing emotional expressiveness in Speech-to-Speech Translation systems. Its innovative methodology and strong experimental results contribute significantly to the field, addressing a critical gap in existing S2ST technologies.
Methodology
The paper introduces a novel synthesis pipeline for creating expressive datasets tailored for Speech-to-Speech Translation (S2ST), addressing the critical issue of data scarcity in capturing non-verbal vocalizations (NVs). The proposed MoVE architecture utilizes a Mixture-of-LoRA-Experts approach, which effectively combines multiple specialized adapters to capture a range of emotional expressions without interference. The dynamic soft-weighting router enhances the model's ability to blend these emotional cues, showcasing a sophisticated understanding of human expressiveness. The methodology is well-structured and innovative, particularly in its approach to leveraging pre-trained AudioLLMs for efficient fine-tuning.
Results & evaluation
The experiments are robust, comparing MoVE against strong baselines and demonstrating superior performance in reproducing NVs and achieving high human-rated naturalness and emotional fidelity. The use of both objective metrics (ASR-BLEU, Aro-Val SIM) and subjective evaluations (Naturalness MOS, Emotion SMOS) provides a comprehensive assessment of the model's capabilities. The results clearly indicate that MoVE outperforms existing systems, establishing a new state-of-the-art in expressive S2ST.
Reproducibility
The paper provides sufficient details regarding the architecture, training process, and evaluation metrics, which should enable other researchers to replicate the experiments. However, the lack of a publicly available code repository may hinder full reproducibility. The authors mention a demo URL, which may provide some insights into the model's performance but does not substitute for complete code access.
Limitations
While the paper makes significant strides in expressive S2ST, it acknowledges that the current approach does not cover all emotional states and NV categories. Additionally, the reliance on pre-trained models may limit the generalizability of the findings to other languages or contexts not represented in the training data.
Broader impact
The advancements presented in this paper could have substantial implications for real-world applications in cross-linguistic communication, enhancing the emotional richness of automated translations. This could improve user experience in various domains, including customer service, entertainment, and virtual communication platforms. The release of the expressive dataset also opens avenues for future research in related fields. This paper presents a pioneering approach to enhancing emotional expressiveness in Speech-to-Speech Translation systems. Its innovative methodology and strong experimental results contribute significantly to the field, addressing a critical gap in existing S2ST technologies.
Author of this paper? Send a correction.