frovolts/Lipi-Ghor-bn-882-SSTT

🗣️ Lipi-Ghor | লিপিঘর — Bengali Speech Dataset (bn-882-SSTT)

   

Lipi-Ghor (লিপিঘর, meaning "House of Scripts") is a large-scale Bengali speech dataset designed for automatic speech recognition (ASR), speaker diarization, and spoken language research. It is one of the largest open Bengali speech corpora with aligned speaker, transcription, and timestamp annotations.

Built by Team_Villagers as part of DL Sprint 4.0.

Dataset Details

Dataset Description

Lipi-Ghor is a large-scale, multi-domain Bengali speech corpus covering ~882 hours of audio sourced from 1,019 YouTube videos across 596 unique channels. Each video has been processed through speaker diarization (pyannote-audio) and aligned with Bengali caption transcripts to produce structured SSTT (Speaker, Speech, Transcription, Timestamp) annotations. The dataset covers a wide range of spoken Bengali domains, registers, and regional dialects, making it one of the most diverse open Bengali speech resources available.

• —Curated by: Team_Villagers — Sanjid Hasan, A H M Fuad, Risalat Labib, Bayazid Hasan
• —Competition: DL Sprint 4.0
• —Language(s): Bengali / Bangla (bn)
• —License: CC BY 4.0

Dataset Sources

• —Repository: Sanjidh090/Lipi-Ghor-bn-882-SSTT
• —Source data: YouTube (public videos with Bengali caption tracks)

Dataset at a Glance

Uses

Direct Use

This dataset is intended for:

• —Bengali ASR model training — fine-tuning Whisper, wav2vec2, MMS, and similar models
• —Speaker diarization research — "who spoke when" tasks in Bengali
• —Bengali TTS — speaker-labeled segments can inform voice synthesis pipelines
• —Dialect identification — the dataset covers Standard Dhaka Bengali, Chittagonian, Sylheti, Rangpuri, and Barishal variants
• —Multilingual NLP benchmarking — Bengali is consistently under-represented in multilingual benchmarks

Out-of-Scope Use

• —Surveillance or speaker re-identification — speaker labels (SPEAKER_00, SPEAKER_01, etc.) are local to each video and do not track identity across videos
• —High-stakes production ASR without filtering — the majority of transcripts are sourced from auto-generated YouTube captions and may contain recognition errors; human verification is recommended before deployment in critical applications

Access this dataset

    from datasets import load_dataset
    
    # Load a sample of Lipi-Ghor
    dataset = load_dataset("Sanjidh090/Lipi-Ghor-bn-882-SSTT", split="test", streaming=True)
    sample = next(iter(dataset))
    
    print(f"Speaker: {sample['speaker']}")
    print(f"Text: {sample['text']}")

Dataset Structure

Lipi-Ghor-bn-882-SSTT/
├── data/                                    # Audio segments (.mp3, pyannote-segmented)
├── diarization_results/                     # Per-video diarization output (*_output.json)
├── diarization_results_with_transcription/  # Diarization + transcript aligned (*_unified.json)
├── diarization_transcription_final/         # Cleaned final outputs (*_unified.json)
└── test/                                    # Test samples (.wav)

File Naming Convention

All annotation files use the YouTube video ID as the base filename:

• —{video_id}_output.json — raw diarization output
• —{video_id}_unified.json — diarization + transcription merged

Annotation Format (SSTT)

Each _unified.json contains an array of segments:

[
  {
    "speaker": "SPEAKER_00",
    "start": 12.34,
    "end": 18.72,
    "text": "আমরা আজকে এই বিষয়টি নিয়ে কথা বলব।"
  }
]

Content & Categories

Dialectal coverage includes: Standard Dhaka Bengali, Chittagonian, Sylheti, Rangpuri, and Barishal variants.

Top Channels by Hours

Dataset Creation

Curation Rationale

Bengali is spoken by over 230 million people yet remains severely under-resourced in ASR and spoken language research. Existing open Bengali ASR datasets are typically small (5–40 hours), limited to read speech, and lack speaker annotations. Lipi-Ghor was created to address this gap with a large-scale, multi-domain, diarized corpus that reflects the diversity of real spoken Bengali across dialects, topics, and recording conditions.

Source Data

Data Collection and Processing

1. 1.Video Selection — YouTube video IDs were collected across 596 Bengali channels covering diverse domains. Only videos with existing Bengali caption tracks (manual or community-contributed) were retained to ensure baseline transcription quality.
2. 2.Audio & Transcript Extraction — `yt-dlp` was used to download audio (MP3) and pull Bengali caption/subtitle tracks (bn language code).
3. 3.Speaker Diarization — `pyannote-audio` was applied to each audio file to segment speech into speaker turns with precise timestamps. Outputs stored as *_output.json.
4. 4.Alignment — YouTube transcripts were aligned with pyannote speaker segments to produce SSTT-format *_unified.json files. A cleaned final version is stored in diarization_transcription_final/.

Who are the Source Data Producers?

The source audio and transcripts are derived from publicly available YouTube content created by Bengali-language content creators across Bangladesh and West Bengal. Content spans professional news channels, independent creators, audiobook narrators, and community contributors.

Annotations

Annotation Process

Speaker diarization was performed automatically using pyannote-audio. Transcription was sourced from existing YouTube caption tracks — 86 videos have manually created captions; the remaining ones use auto-generated YouTube captions. Alignment between diarization segments and caption timestamps was performed programmatically.

Who are the Annotators?

Diarization: pyannote-audio (automated). Transcription: YouTube caption system and original content creators. Post-processing and pipeline: Team_Villagers (Sanjid Hasan, A H M Fuad, Risalat Labib, Bayazid Hasan).

Personal and Sensitive Information

The dataset contains publicly broadcast speech from YouTube. Speaker labels are anonymous (SPEAKER_00, etc.) and are not linked to real-world identities. No cross-video speaker identity tracking is performed. Content creators retain their original copyright; this dataset is intended for research and non-commercial use only.

If you are a content creator and wish to have your content removed from this dataset, please open an issue or contact us directly.

Bias, Risks, and Limitations

• —Transcript quality varies — 86 videos have human-verified captions; 1,254 use auto-generated YouTube captions which may contain recognition errors, especially for dialectal speech and code-switching.
• —Audio quality varies — sourced from diverse YouTube content; some recordings contain background music, overlapping speakers, or artifacts.
• —~194 hours pending — approximately 321 videos are sourced and diarized but not yet fully uploaded to this repository.
• —Speaker labels are local — SPEAKER_00, SPEAKER_01 etc. are per-video labels only. Cross-video speaker identity is not tracked.
• —Code-switching — some content contains Bengali-English mixing, which reflects real usage but may affect monolingual ASR models.
• —Geographic bias — the majority of content originates from Dhaka-centric media channels; rural and minority dialects may be underrepresented relative to their speaker populations.

Recommendations

Users training ASR models should consider filtering by transcript type (manual vs auto) and evaluating on a held-out human-verified subset before deployment. For dialect-robust training, stratified sampling across the category and channel distribution is recommended.

Citation

If you use Lipi-Ghor in your research, please cite:

BibTeX:

@dataset{lipighor2026,
  title     = {Lipi-Ghor: A Large-Scale Bengali Speech Dataset with Speaker Diarization and Transcription},
  author    = {Hasan, Sanjid and Fuad, A. H. M. and Labib, Risalat and Hasan, Bayazid},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/7877},
  url       = {https://huggingface.co/datasets/Sanjidh090/Lipi-Ghor-bn-882-SSTT},
  note      = {DL Sprint 4.0, Team Villagers}
}

APA:

Hasan, S., Fuad, A. H. M., Labib, R., & Hasan, B. (2025). Lipi-Ghor: A Large-Scale Bengali Speech Dataset with Speaker Diarization and Transcription [Dataset]. HuggingFace. https://huggingface.co/datasets/Sanjidh090/Lipi-Ghor-bn-882-SSTT

Paper Citation:

@misc{hasan2026make,
  title         = {Make It Hard to Hear, Easy to Learn: Long-Form Bengali ASR and Speaker Diarization via Extreme Augmentation and Perfect Alignment},
  author        = {Hasan, Sanjid and Labib, Risalat and Fuad, A. H. M. and Hasan, Bayazid},
  year          = {2026},
  eprint        = {2602.23070},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2602.23070}
}

Glossary

Dataset Card Authors

Team_Villagers — DL Sprint 4.0

• —Sanjid Hasan
• —A H M Fuad
• —Risalat Labib
• —Bayazid Hasan

Dataset Card Contact

Open an issue on the HuggingFace repository or contact via the repository discussion tab.

Acknowledgements

• —yt-dlp for audio and caption extraction
• —pyannote-audio for speaker diarization
• —All Bengali content creators whose work made this dataset possible
• —DL Sprint 4.0 organizers

লিপিঘর — বাংলা ভাষার জন্য, বাংলা ভাষার গবেষকদের জন্য। Lipi-Ghor — for the Bengali language, for Bengali language researchers.