Leveraging Audio-LLMs to Filter Speech-to-Speech Training Data
Researchers have developed a novel method using audio-large language models (Audio-LLMs) to filter noisy speech-to-speech translation (S2ST) training data. This approach employs a two-stage Rank-to-Distill strategy, where an initial ranker generates pseudo-labels for keeping or dropping speech pairs, which then train an Audio-LLM to make these decisions directly from audio. The model effectively captures acoustic fidelity and cross-lingual semantic consistency, leading to significant improvements in S2ST performance, with gains of up to +1.4 ASR-BLEU on benchmark datasets. AI
IMPACT Improves the quality of training data for speech translation models, potentially leading to more accurate and robust speech-to-speech translation systems.