smam/bengali-diarization-synthetic-v4

Bengali Speaker Diarization Synthetic Dataset V4

Synthetic Bengali speaker diarization dataset with natural overlapping speech patterns using timeline-based random chunk placement.

Dataset Overview

Property	Value
Total Samples	600
Speaker Categories	1-30 speakers per sample
Samples per Category	20
Duration per Sample	~30 minutes
Total Duration	~300 hours
Sample Rate	16000 Hz
Format	WAV (audio) + RTTM (labels) + JSON (metadata)

Generation Method

Simple Timeline-Based Approach:

1. Select k speakers from 324 Bengali voices
2. For each speaker, place audio chunks at random positions on a 30-minute timeline
3. Chunk durations follow lognormal distribution: μ=log(25), σ=0.8 → mostly 10-60s, up to 300s
4. Overlaps emerge naturally from independent placement (no artificial overlap injection)

This creates realistic overlapping speech patterns where speakers naturally talk over each other.

Structure

{k}_speaker/
├── simple_{k}spk_000.wav   # Audio file
├── simple_{k}spk_000.rttm  # RTTM labels
├── simple_{k}spk_000.json  # Metadata
└── ...

RTTM Format

SPEAKER <file_id> 1 <start> <duration> <NA> <NA> <speaker_label> <NA> <NA>

Usage

from datasets import load_dataset
ds = load_dataset("smam/bengali-diarization-synthetic-v4")

# Or download specific folders
from huggingface_hub import snapshot_download
snapshot_download("smam/bengali-diarization-synthetic-v4", 
                  allow_patterns="5_speaker/*",
                  local_dir="./data")

Source Data

Generated from:

• IndicVoices (Bengali subset)
• Kathbath dataset (Bengali)

Total: 324 qualified Bengali speakers with ≥1 minute of audio each.

License

CC-BY-4.0