Scaling Conversational Hungarian ASR: The BEA-Dialogue+ Corpus
Researchers have introduced BEA-Dialogue+, an expanded corpus for Hungarian conversational automatic speech recognition (ASR). This new dataset increases the available training data to 200 hours, relaxing split criteria to allow for more material while maintaining speaker separation. Evaluations using Whisper and FastConformer models demonstrate that the larger dataset, especially when combined with Serialized Output Training (SOT) fine-tuning, leads to significant improvements in transcription accuracy metrics. AI
IMPACT Provides a larger, more challenging benchmark for Hungarian dialogue ASR, enabling better training and evaluation of transcription systems.