NbAiLab/nb-asr-qwen3whisperxagreement-v1
收藏资源简介:
该数据集是挪威语语音的词级强制对齐训练数据,通过保留两个独立对齐器(WhisperX和Qwen3)在严格容忍度内达成一致的示例来创建。数据集包含702,067个从NB-ASR挪威语音频语料库中提取的语音片段,每个记录将音频文件与词级强制对齐配对,格式适用于训练预测时间戳(如<timestamp>)的序列模型(例如Qwen3微调)。时间戳以80毫秒分辨率的整数帧索引表示(12.5帧/秒),与对齐过程中使用的特征帧率匹配。数据创建过程涉及运行两个强制对齐系统,并比较词边界:仅当每个词的开始时间差不超过2帧(≤160毫秒)且结束时间差不超过5帧(≤400毫秒)时,才保留对齐对;如果两个系统的词数不同,则直接拒绝样本。数据集记录为JSON对象,包含id(唯一片段标识符)、audio(16 kHz单声道WAV文件的绝对路径)、text(带有<timestamp>占位符标记的词序列)和labels(帧索引标签,词位置为-100,<timestamp>位置为开始帧和结束帧)。数据集旨在用于监督微调从音频预测词级时间戳的模型,双重一致过滤器提供比单独使用任一对齐器更高置信度的地面真实边界。源音频来自NB-ASR挪威语广播和议会语音语料库,文件为16 kHz单声道WAV格式。
This dataset contains word-level forced alignment training data for Norwegian speech, produced by keeping only examples where two independent aligners — WhisperX and Qwen3 (Lunde forced aligner) — agree within a tight tolerance. It includes 702,067 speech segments drawn from the NB-ASR Norwegian audio corpus. Each record pairs an audio file with a word-level forced alignment in a format suitable for training a <timestamp>-predicting sequence model (e.g., Qwen3 fine-tuning). Timestamps are represented as integer frame indices at 80 ms resolution (12.5 frames/sec), matching the feature-frame rate used during alignment. The data creation process involved running two forced-alignment systems and comparing word boundaries: alignments are kept only when every word satisfies a start tolerance of ≤2 frames (≤160 ms) and an end tolerance of ≤5 frames (≤400 ms); samples with different word counts are rejected outright. Each record is a JSON object with fields: id (unique segment identifier), audio (absolute path to 16 kHz mono WAV file), text (words interleaved with <timestamp> placeholder tokens), and labels (frame-index labels with -100 at word positions and start/end frames at <timestamp> positions). The dataset is intended for supervised fine-tuning of models that predict word-level timestamps from audio, with the double-agreement filter providing higher-confidence ground-truth boundaries. Source audio originates from the NB-ASR Norwegian broadcast and parliamentary speech corpus, stored as 16 kHz mono WAV files.




