nymtheescobar/bengali-talkshow-audio

Bengali Talkshow Audio Dataset

A large-scale collection of 1,180 Bengali talk show audio recordings totaling 789+ hours of multi-speaker speech, sourced from Bangladeshi television talk shows and political debate programs.

Dataset Description

This dataset contains audio from Bengali-language TV talk shows, political debates, and news discussion programs from major Bangladeshi television channels. Each recording features multiple speakers engaged in discussion, making it ideal for:

• —Speaker Diarization ("who spoke when?")
• —Bengali Automatic Speech Recognition (ASR)
• —Speaker Recognition / Verification
• —Voice Activity Detection (VAD)
• —Bengali Language Modeling
• —Multi-speaker audio analysis

Dataset Summary

Duration Distribution

Top Source Channels

How to Use

Quick Start

from datasets import load_dataset

ds = load_dataset("nymtheescobar/bengali-talkshow-audio", split="train")
print(f"Total recordings: {len(ds)}")

# Access a sample
sample = ds[0]
print(sample["title"])
print(sample["channel"])
print(f"Duration: {sample['duration_min']} min")

Load with Audio

from datasets import load_dataset, Audio

ds = load_dataset("nymtheescobar/bengali-talkshow-audio", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

# Now audio is decoded on-the-fly
sample = ds[0]
waveform = sample["audio"]["array"]       # numpy array
sr = sample["audio"]["sampling_rate"]       # 16000

Stream (No Full Download)

from datasets import load_dataset, Audio

ds = load_dataset("nymtheescobar/bengali-talkshow-audio", split="train", streaming=True)
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

for sample in ds:
    print(sample["title"], sample["duration_min"], "min")
    audio = sample["audio"]["array"]
    break  # just peek at first

Filter by Duration / Channel

# Only recordings > 30 min
long_shows = ds.filter(lambda x: x["duration_min"] >= 30)

# Only ATN Bangla
atn = ds.filter(lambda x: "ATN" in (x["channel"] or ""))

Dataset Structure

Data Fields

Audio Specifications

The 16kHz mono format is the standard for speech processing and is directly compatible with:

• —Whisper (OpenAI)
• —Wav2Vec 2.0 / HuBERT (Meta)
• —pyannote.audio
• —SpeechBrain
• —NeMo (NVIDIA)
• —IndicWav2Vec

Use Cases

Speaker Diarization

Each recording contains multiple speakers (typically 3-20+) in conversational settings with natural turn-taking, overlapping speech, and varying acoustic conditions. This makes the dataset particularly valuable for developing and benchmarking Bengali speaker diarization systems.

Bengali ASR

With ~789 hours of Bengali speech from diverse speakers across 228 channels, this dataset can supplement existing Bengali ASR training data or be used for domain adaptation.

Speaker Recognition

The dataset covers 228+ unique TV channels with recurring speakers (hosts, politicians, analysts), enabling Bengali speaker recognition and verification research.

Source and Collection

Citation

If you use this dataset in your research, please cite:

@dataset{islam2026bengali_talkshow_audio,
  title={Bengali Talkshow Audio Dataset},
  author={Islam, Naymul},
  year={2026},
  url={https://huggingface.co/datasets/nymtheescobar/bengali-talkshow-audio},
  note={1,180 recordings, ~789 hours of Bengali multi-speaker speech from TV talk shows}
}

License

This dataset is released under CC-BY-4.0. The original audio content is sourced from publicly available YouTube videos. Please respect the original content creators' rights.

Author

Naymul Islam

• —HuggingFace: nymtheescobar
• —Email: naymul504@gmail.com

Acknowledgments

• —Audio sourced from publicly available Bangladeshi television content on YouTube
• —Built as part of BUET DL Sprint 4.0 (BUET CSE Fest 2026) for Bengali Speaker Diarization research