Nepali–English Code-Switched ASR 是一个用于自动语音识别(ASR)模型训练与评估的语码转换语音数据集。该数据集专注于尼泊尔语-英语混合语音,这种语言变体在尼泊尔高等教育和在线辅导中普遍使用,教师通常在尼泊尔语语法中流利地混入英语技术词汇。数据集内容源自公开的YouTube STEM(科学、技术、工程、数学)和计算机科学讲座视频,涵盖了算法、数据结构、电子学、物理和数学等主题。原始视频经过语音活动检测分割成片段,并排除了前90秒可能含有片头音乐的部分。音频经过响度归一化处理。数据规模总计约58.7小时,包含20,886个音频片段。音频格式为16 kHz、单声道、16位的WAV文件。数据已划分为训练集(16,868个片段,占80.8%)、验证集(2,281个片段,占10.9%)和测试集(1,737个片段,占8.3%)。根据语音内容,片段被标注为三种类型:语码转换(`cs`,19,670个片段)、以英语为主(`en`,651个片段)和以尼泊尔语为主(`ne`,565个片段)。数据来源于20个不同的YouTube频道,共163个视频。数据集包含以下字段:`audio`(响度归一化的WAV音频片段)、`video`(源MP4视频片段)、`transcription`(混合脚本的规范转录文本,遵循“罗马化尼泊尔英语”惯例)、`transcription_romanized`(拉丁字母音译版本)、`speech_type`(语音类型)、`cmi`(语码混合指数)、`speaker_id`(说话人ID)、`duration_s`(片段时长)。数据集主要用于语码转换自动语音识别任务,通过语码混合指数(CMI)字段,用户可以筛选不同混合程度的语音片段进行模型训练或分析。
Nepali–English Code-Switched ASR is a code-switched speech dataset for automatic speech recognition (ASR) model training and evaluation. It focuses on Nepali-English mixed speech, a language variant commonly used in higher education and online tutoring in Nepal, where teachers fluently incorporate English technical terms within Nepali grammar. The dataset is derived from publicly available YouTube STEM (Science, Technology, Engineering, Mathematics) and computer science lecture videos, covering topics such as algorithms, data structures, electronics, physics, and mathematics. Original videos were segmented using voice activity detection, excluding the first 90 seconds to avoid potential intro music. Audio underwent loudness normalization. The total data size is approximately 58.7 hours, comprising 20,886 audio segments. Audio files are in 16 kHz, mono, 16-bit WAV format. The data is split into training set (16,868 segments, 80.8%), validation set (2,281 segments, 10.9%), and test set (1,737 segments, 8.3%). Based on speech content, segments are labeled into three types: code-switched (`cs`, 19,670 segments), English-dominant (`en`, 651 segments), and Nepali-dominant (`ne`, 565 segments). Data originates from 20 different YouTube channels, totaling 163 videos. The dataset includes the following fields: `audio` (loudness-normalized WAV audio segment), `video` (source MP4 video segment), `transcription` (canonical transcription text in mixed script, following Romanized Nepali English conventions), `transcription_romanized` (Latin script transliteration of the entire transcription), `speech_type` (speech type), `cmi` (code-mixing index), `speaker_id` (speaker ID, i.e., YouTube channel ID), and `duration_s` (segment duration in seconds). The dataset is primarily used for code-switched ASR tasks, and users can filter segments by different mixing levels using the code-mixing index (CMI) field for model training or analysis.