The article discusses a novel approach to developing a speech-to-text (S2T) data engine (DE) utilizing a text-only language model (LLM). The research demonstrates that this method can outperform previous S2T systems, achieving better outcomes with significantly reduced training data. Furthermore, the authors highlight the potential for enhancing zero-shot speech translation by integrating existing translation datasets with S2T data. Their findings span across 102 languages for speech-to-text and 4 for speech-to-text translation products, emphasizing broadened applications in cross-lingual scenarios.
We present an effective approach to developing a speech-to-text DE from a text-only LLM, suggesting that it can outperform previous methods with less training data.
Our findings indicate that zero-shot speech translation can be improved by combining readily available translation and speech-to-text data, showcasing that innovation can be driven by data synergy.
#speech-to-text #language-models #deep-learning #cross-lingual-processing #natural-language-processing
Collection
[
|
...
]