anhndbk/ViASRQuantizationBench

ViASR-Bench是一个越南语自动语音识别（ASR）模型量化的基准数据集，提供校准集和评估集。该数据集是从BUD500（最大的越南语播客语音公开数据集）中精心挑选的样本，专门设计用于ASR量化的两阶段生命周期。越南语具有六声调系统和多种方言变体，这使得使用英语语音数据进行校准会导致量化效果不佳。ViASR-Bench通过提供与目标语言声学分布匹配的校准数据集，解决了这一问题。数据集包含256个校准话语和600个评估话语，音频格式为16 kHz单声道PCM-16（WAV）。此外，数据集还提供了详细的使用方法、质量过滤条件和采样方法，确保数据的多样性和非重叠性。

ViASR-Bench is a Vietnamese benchmark dataset for ASR model quantization, providing both a calibration set and an evaluation set. It is a curated sample from BUD500, the largest public Vietnamese podcast speech dataset, and is designed specifically for the two-phase lifecycle of ASR quantization. Vietnamese has a six-tone tonal system and significant regional dialect variation, which makes calibrating on English audio suboptimal. ViASR-Bench addresses this by providing a calibration dataset whose acoustic distribution matches the target language. The dataset includes 256 calibration utterances and 600 evaluation utterances, with audio in 16 kHz mono PCM-16 (WAV) format. It also provides detailed usage instructions, quality filters, and sampling methodology to ensure diversity and non-overlapping splits.