Efficient ASR Training with Conversations that Never Happened
Researchers have developed a novel method to enhance Automatic Speech Recognition (ASR) training for low-resource languages by generating synthetic conversational data. This pipeline uses LLMs to create dialogues, maps speaker attributes to TTS voice profiles, and assembles simulated conversations. Evaluations on the Hungarian BEA-Dialogue benchmark showed that this synthetic data significantly improves ASR performance, even outperforming models trained on much larger real datasets. AI
IMPACT Synthetic data generation via LLMs and TTS offers a scalable solution for improving ASR in low-resource languages.