Dataset Card for SUBAK.KO

Dataset Summary

SUBAK.KO (সুবাক্য), a publicly available annotated Bangladeshi standard Bangla speech corpus, is compiled for automatic speech recognition research. This corpus contains 241 hours of high-quality speech data, including 229 hours of read speech data and 12 hours of broadcast speech data. The read speech segment is recorded in a noise-proof studio environment from 33 male and 28 female native Bangladeshi Bangla speakers representing 8 divisions/34 districts of Bangladesh. Furthermore, the read speech segment comprises a total of 1 hour and 30 minutes of recorded speech provided by two second language (L2) speakers. The broadcast speech segment is collected from YouTube. SUBAK.KO has been manually annotated under human supervision to ensure gold-standard labels. The corresponding paper reports detailed information about the development and baseline performance of SUBAK.KO and cross-dataset evaluation in comparison to LB-ASRTD corpus.

SUBAK.KO is developed by the researchers from the Department of Computer Science and Engineering (CSE) at Shahjalal University of Science and Technology (SUST), Bangladesh with financial support from the Higher Education Quality Enhancement Project (AIF Window 4, CP 3888) for “The Development of Multi-Platform Speech and Language Processing Software for Bangla” of the University Grants Commission (UGC), Bangladesh.

Example Usage

To load the full SUBAK.KO corpus, use the following code:

from datasets import load_dataset

dataset = load_dataset("SUST-CSE-Speech/SUBAK.KO")

To load a specific split of the SUBAK.KO, define the split and set the streaming mode as True in the following way:

from datasets import load_dataset

dataset = load_dataset("SUST-CSE-Speech/SUBAK.KO", split="test", streaming=True)

More documentation on streaming can be found from this link.

Alternatively, you can manually download the zipped SUBAK.KO folder from this HuggingFace directory. The csv files corresponding to the train, validation and test splits can be found in the same directory.

Supported Tasks and Leaderboards

This dataset is designed for the automatic speech recognition task. The associated paper provides the baseline results on SUBAK.KO corpus.

Languages

Bangladeshi standard Bangla

Dataset Structure

Data Instances

A typical data point comprises the path to the audio file and its transcription.

{
 'audio': {'path': '/home/username/subakko/part5/wav5/e4/TNM22_MESBA_page_257-258_5_5_Labeled_by_Tomal-20.wav',
           'array': array([-0.00048828, -0.00018311, -0.00137329, ...,  0.00079346, 0.00091553,  0.00085449], dtype=float32),
           'sampling_rate': 16000},
 'transcript': 'তারপর চার মাস তিনি ছিলেন কেন্দ্রীয় গোয়েন্দা সংস্থার তত্বাবধানে এক নিরাপদ জায়গায়',
 'path': '/subakko/part5/wav5/e4/TNM22_MESBA_page_257-258_5_5_Labeled_by_Tomal-20.wav'
}

Data Fields

• audio: A dictionary containing the path to the original audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0].

• transcription: The orthographic transcription

• file_path: The relative path to the audio file

Data Splits

SUBAK.KO has been subdivided into three splits for train, validation and test. It is strongly advised to use identical data splits for research purposes to facilitate benchmarking across various models.

	Train	Validation	Test
Utterances	64491	6594	6533
Duration	200.3 hrs	20.5 hrs	20.3 hrs

Additional Information

Licensing Information

CC BY 4.0

Citation Information

Please cite the following paper if you use the corpus.

@article{kibria2022bangladeshi,
  title={Bangladeshi Bangla speech corpus for automatic speech recognition research},
  author={Kibria, Shafkat and Samin, Ahnaf Mozib and Kobir, M Humayon and Rahman, M Shahidur and Selim, M Reza and Iqbal, M Zafar},
  journal={Speech Communication},
  volume={136},
  pages={84--97},
  year={2022},
  publisher={Elsevier}
}

Contributions

Thanks to Ahnaf Mozib Samin for adding this dataset.