Suprio85/Bangla_speech_corpus-321
šļø BanglaSpeechCorpus-321: Large-Scale Long-Form Bangla Speech Corpus Dataset Summary BanglaSpeechCorpus-321 is an extended, large-scale Bangla (Bengali) speech corpus for Automatic Speech Recognition (ASR), featuring 321.2 hours of naturally occurring Bangla speech across 401 recordings. This is the expanded successor to Bangla_Speech_Corpus, covering a broader set of YouTube channels including drama serials, audiobooks, and entertainment content. With over 303ā¦ See the full description on the dataset page: https://huggingface.co/datasets/Suprio85/Bangla_speech_corpus-321.
šļø BanglaSpeechCorpus-321: Large-Scale Long-Form Bangla Speech Corpus
Dataset Summary
BanglaSpeechCorpus-321 is an extended, large-scale Bangla (Bengali) speech corpus for Automatic Speech Recognition (ASR), featuring 321.2 hours of naturally occurring Bangla speech across 401 recordings. This is the expanded successor to Bangla_Speech_Corpus, covering a broader set of YouTube channels including drama serials, audiobooks, and entertainment content.
With over 303,000 segments, 1.79 million words, and a maximum segment duration of 20 seconds, this corpus is purpose-built for long-form, continuous-transcript ASR evaluation and training ā a regime where current state-of-the-art models remain significantly undertested for Bangla.
Key Statistics
Source Channels
Channels include all sources from the original BanglaVoice-LF 191 corpus, plus additional channels:
Full per-channel breakdown is available in video_checklist.csv.Dataset Structure
Bangla_speech_corpus-321/
āāā audio/ # 401 WAV audio files (16 kHz, mono)
āāā subtitles_raw/ # Raw YouTube subtitle files
āāā transcripts/ # Cleaned, aligned transcript JSONs (401 files)
āāā video_checklist.csv # Per-video metadata
āāā completed_videos.log # Processing logmanifest.jsonl Schema
Each line is a JSON object representing one segment:
{
"id": "video_id_seg_00042",
"audio_path": "audio/video_id.wav",
"transcript": "ą¦¬ą¦¾ą¦ą¦²ą¦¾ ą¦ą§ą¦°ą¦¾ą¦Øą§ą¦øą¦ą§ą¦°ą¦æą¦Ŗą§ą¦ ą¦ą¦ą¦¾ą¦Øą§ ą¦²ą§ą¦ą¦¾ ą¦ą¦ą§",
"channel": "Banglavision DRAMA",
"start": 124.5,
"end": 139.2,
"duration": 14.7,
"source_url": "https://www.youtube.com/watch?v=..."
}Transcript JSON Schema
Each file in transcripts/ contains:
{
"video_id": "abc123",
"channel": "Eagle Premier Station",
"duration_seconds": 2910.0,
"segments": [
{ "start": 0.0, "end": 18.4, "text": "..." },
{ "start": 18.4, "end": 35.1, "text": "..." }
]
}How to Use
Basic Loading
from datasets import load_dataset
dataset = load_dataset("Suprio85/Bangla_speech_corpus-321")
print(dataset)Load with Audio Column
from datasets import load_dataset, Audio
dataset = load_dataset("Suprio85/Bangla_speech_corpus-321", split="train")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
sample = dataset[0]
print("Transcript:", sample["transcript"])
print("Duration: ", sample["duration"], "seconds")
print("Channel: ", sample["channel"])Stream Large Dataset (Recommended for 321h)
from datasets import load_dataset
# Streaming avoids downloading all 206 GB upfront
dataset = load_dataset(
"Suprio85/Bangla_speech_corpus-321",
split="train",
streaming=True
)
for sample in dataset.take(5):
print(sample["transcript"])Fine-tuning Whisper on This Dataset
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset, Audio
processor = WhisperProcessor.from_pretrained("openai/whisper-large-v3")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v3")
model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(
language="bengali", task="transcribe"
)
dataset = load_dataset("Suprio85/Bangla_speech_corpus-321", split="train")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
def prepare_batch(batch):
audio = batch["audio"]
batch["input_features"] = processor(
audio["array"],
sampling_rate=audio["sampling_rate"],
return_tensors="pt"
).input_features[0]
batch["labels"] = processor.tokenizer(batch["transcript"]).input_ids
return batch
dataset = dataset.map(prepare_batch, remove_columns=dataset.column_names)Comparison: 155h vs 321h
Motivation
Long-form continuous-transcript evaluation in Bangla ASR remains severely limited. Most public datasets consist of short, isolated, studio-recorded utterances that do not reflect real-world usage. BanglaVoice-LF 321 directly addresses this by providing:
- 321 hours of naturalistic in-the-wild Bangla speech
- Continuous, full-episode transcripts from broadcast drama and entertainment
- 20-second max segments suitable for both training and evaluation pipelines
- 1.79M words of rich, colloquial, and formal Bangla vocabulary
Intended Uses
- Training and fine-tuning Bangla ASR models (Whisper, wav2vec2, MMS)
- Long-form speech recognition benchmarking
- Bangla language model and tokenizer training
- Speech segmentation and forced alignment research
- Downstream Bangla NLP tasks on transcript text
Limitations
- Transcripts are derived from YouTube auto-subtitles and may contain errors in emotionally expressive or fast speech.
- Background music and overlapping speech are present in some drama recordings.
- Speaker-level metadata (gender, age, region) is not annotated.
Citation
If you use this dataset in your research, please cite:
@dataset{Bangla_Speech_Corpus-321_2025,
author = {Suprio85},
title = {BanglaVoice-LF 321: Large-Scale Long-Form Bangla Speech Corpus},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/Suprio85/Bangla_speech_corpus-321}
}Also consider citing the original 155h corpus:
@dataset{Bangla_Speech_Corpus_2025,
author = {Suprio85},
title = {BanglaVoice-LF: A Long-Form Bangla Speech Corpus},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/Suprio85/Bangla_Speech_Corpus}
}License
Released under Creative Commons Attribution 4.0 International (CC BY 4.0). Free to use, share, and adapt with attribution.
Related Datasets
- š Bangla_Speech_Corpus ā The original 155-hour corpus