Dataset Card for BanSpeech

Table of Contents

• Dataset Card for SUBAK.KO
  • Table of Contents
  • Dataset Description
    • Dataset Summary
    • Supported Tasks and Leaderboards
    • Languages
  • Dataset Structure
    • Data Instances
    • Data Fields
    • Data Splits
  • Dataset Creation
    • Curation Rationale
    • Source Data
      • Initial Data Collection and Normalization
      • Who are the source language producers?
    • Annotations
      • Annotation process
      • Who are the annotators?
    • Personal and Sensitive Information
  • Considerations for Using the Data
    • Social Impact of Dataset
    • Discussion of Biases
    • Other Known Limitations
  • Additional Information
    • Dataset Curators
    • Licensing Information
    • Citation Information
    • Contributions

Dataset Description

• Developed By Dept. of CSE, SUST, Bangladesh
• Paper: BanSpeech: A Multi-domain Bangla Speech Recognition Benchmark Toward Robust Performance in Challenging Conditions
• Point of Contact: Ahnaf Mozib Samin

Dataset Summary

BanSpeech is a publicly available human-annotated Bangladeshi standard Bangla multi-domain automatic speech recognition (ASR) benchmark. This benchmark contains approximately 6.52 hours of human-annotated broadcast speech, totaling 8085 utterances, across 13 distinct domains and is primarily designed for ASR performance evaluation in challenging conditions e.g. spontaneous, domain-shifting, multi-talker, code-switching. In addition, BanSpeech covers dialectal domains from 7 regions of Bangladesh, however, this part is weakly labeled and can be used for dialect recognition task. The corresponding paper reports detailed information about the development of BanSpeech, along with an analysis of the performance of state-of-the-art fully supervised, self-supervised, and weakly supervised models on BanSpeech.

BanSpeech is developed by the researchers from the Department of Computer Science and Engineering (CSE) at Shahjalal University of Science and Technology (SUST), Bangladesh.

Example Usage

To load the full BanSpeech, use the following code:

from datasets import load_dataset

dataset = load_dataset("SUST-CSE-Speech/banspeech")

To load a specific domain of the BanSpeech, define the domain in the split parameter and set the streaming mode as True in the following way:

from datasets import load_dataset

dataset = load_dataset("SUST-CSE-Speech/banspeech", split="sports", streaming=True)

More documentation on streaming can be found from this link.

Alternatively, you can manually download the BanSpeech from this HuggingFace directory. The compressed folder contains speeches from the 13 general domains as well as the 7 dialectal domains. The csv files corresponding to the domains can be found in the same zipped file.

Supported Tasks and Leaderboards

This benchmark is designed for the automatic speech recognition performance evaluation. The associated paper provides the comprehensive evaluation of the state-of-the-art models on BanSpeech.

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/Study/wav2vec2/bangla_broadcast_speech_corpus/banspeech/television_news/news_shomoy_11_d_222.wav',
		   'array': array([-0.00048828, -0.00018311, -0.00137329, ...,  0.00079346, 0.00091553,  0.00085449], dtype=float32),
		   'sampling_rate': 16000},
 'transcript': 'এবং রাস্তা হয়েছে',
 'path': '/television_news/news_shomoy_11_d_222.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

Additional Information

Licensing Information

CC BY 4.0

Citation Information

Please cite the following paper if you use the corpus.

@ARTICLE{10453554,
  author={Samin, Ahnaf Mozib and Kobir, M. Humayon and Rafee, Md. Mushtaq Shahriyar and Ahmed, M. Firoz and Hasan, Mehedi and Ghosh, Partha and Kibria, Shafkat and Rahman, M. Shahidur},
  journal={IEEE Access}, 
  title={BanSpeech: A Multi-Domain Bangla Speech Recognition Benchmark Toward Robust Performance in Challenging Conditions}, 
  year={2024},
  volume={12},
  number={},
  pages={34527-34538},
  keywords={Speech recognition;Data models;Benchmark testing;Speech processing;Robustness;Solid modeling;Task analysis;Automatic speech recognition;Transfer learning;Neural networks;Convolutional neural networks;Supervised learning;Automatic speech recognition;Bangla;domain shifting;read speech;spontaneous speech;transfer learning},
  doi={10.1109/ACCESS.2024.3371478}}

Contributions

Thanks to Ahnaf Mozib Samin for adding this dataset.