Speech Recognition Datasets for Congolese Languages

This dataset contains two new benchmark corpora designed for low-resource languages spoken in the Democratic Republic of the Congo: The Lingala Read Speech Corpus LRSC, with 4.3 hours of labelled audio, and the Congolese Speech Radio Corpus CSRC, which offers 741 hours of unlabeled audio spanning four significant low-resource languages of the region (Lingala, Tshiluba, Kikongo and Congolese Swahili). Collecting speech and audio for this dataset involved two sets of processes: (1) for LRSC, 32 Congolese adult participants were instructed to sit in a relaxed manner within centimetres of an audio recording device or smartphone and read from the text utterances; (2) for CSRC, recording from the archives of a broadcast station were pre-processed and curated. Congolese languages tend to fall into the “low-resource” category, which, in contrast to “high-resource” languages, has fewer datasets accessible, limiting the development of Conversational Artificial Intelligence. This results in creating the speech recognition datasets for low-resource Congolese languages. The proposed dataset contains two sections. The first section involves training a supervised speech recognition module, while the second involves pre-training a self-supervised model. Both sections feature a wide variety of speech and audio taken in various environments, with the first section featuring a speech having its corresponding transcription and the second featuring a collection of pre-processed raw audio data.

本数据集包含两套专为刚果民主共和国境内使用的低资源语言打造的新型基准语料库：林加拉朗读语音语料库（Lingala Read Speech Corpus, LRSC），包含4.3小时标注音频；以及刚果语音广播语料库（Congolese Speech Radio Corpus, CSRC），涵盖741小时未标注音频，覆盖该地区四种重要低资源语言：林加拉语、奇卢伯语、吉孔戈语与刚果斯瓦希里语。

本数据集的语音与音频采集包含两套流程：（1）针对LRSC，招募32名刚果成年参与者，令其保持放松坐姿，在距离音频录制设备或智能手机数厘米范围内朗读文本语句；（2）针对CSRC，对某广播电台存档的录音进行预处理与精选整理。

刚果语种大多归入“低资源”类别，相较于“高资源”语言，其可获取的数据集极为有限，制约了会话式人工智能（Conversational Artificial Intelligence）的发展。正因如此，本数据集应运而生，为刚果低资源语言搭建了专属的语音识别语料。

本数据集包含两个数据模块：第一模块用于训练有监督语音识别模型，第二模块用于自监督模型的预训练。两类模块均涵盖多场景下采集的多样化语音与音频数据，其中第一模块附带对应转录文本的语音数据，第二模块则包含预处理后的原始音频数据集。