IntisarUddin/Bengali_Long_form_ASR

Bengali Long-Form ASR Dataset

Dataset Summary

The Bengali Long-Form ASR Dataset is a large-scale collection of long-duration Bangla speech recordings paired with verified transcripts. The dataset is designed specifically for long-form Automatic Speech Recognition (ASR) research.

Key Statistics

• Total duration: 310.06 hours
• Number of recordings: 382
• Average duration per recording: ~48.7 minutes
• Language: Bengali (bn)
• Audio format: WAV
• Sampling rate: 16 kHz
• Channels: Mono

This dataset is suitable for:

• Long-form ASR research
• Whisper fine-tuning
• Transformer-based speech modeling
• Low-resource speech research

---

Dataset Structure

root/
├── audio/
├── dataset_metadata.csv
└── dataset_metadata.json

---

Audio

All recordings are stored in the audio/ directory.

Example: audio/audio_001.wav

Audio Properties:

• Format: WAV
• Sampling rate: 16 kHz
• Mono channel

---

Metadata

The metadata file (dataset_metadata.csv) contains:

• file_name: Audio file name (e.g., audio_001.wav)
• original_id: Original YouTube video ID
• title: Original video title
• duration_min: Duration of recording in minutes
• transcription: Full transcription text

---

Transcription Method

Transcriptions were generated using YouTube auto-captions or creator-provided subtitles, followed by manual correction and normalization.

This semi-automatic approach balances scalability and transcription quality.

---

Dataset Splitting

This dataset does NOT include predefined train/validation/test splits.

Since speaker metadata is not available, users are advised to perform recording-wise splitting to avoid data leakage.

Recommended split strategy:

• Train: 80%
• Validation: 10%
• Test: 10%

Splitting should be performed at the audio file level, not at the text level.

---

Intended Use

This dataset is intended for academic and research use in:

• Long-form Automatic Speech Recognition
• Context-aware ASR systems
• End-to-end speech modeling
• Bangla speech-language modeling

---

Known Limitations

• No speaker annotations
• No predefined dataset splits
• Long-form audio requires segmentation before training
• Possible domain imbalance
• Minor transcription inconsistencies may remain

---

Loading the Dataset

Using Hugging Face

from datasets import load_dataset

dataset = load_dataset("IntisarUddin/Bengali_Long_form_ASR")

print(dataset)

---

Or Manually Using Pandas

import pandas as pd

df = pd.read_csv("dataset_metadata.csv")

print(df.head())

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY 4.0) License.

Users are free to share and adapt the material for any purpose, even commercially, provided appropriate credit is given.

---

Citation

@dataset{IntisarBengaliLongFormASR2026, author = {K M Intisar Uddin}, title = {Bengali Long-Form ASR Dataset}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/datasets/IntisarUddin/Bengali_Long_form_ASR }, license = {CC-BY 4.0} }