# DL Sprint 4.0 | Bengali Speaker Diarization — Data

来源：https://www.kaggle.com/competitions/dl-sprint-4-0-bengali-speaker-diarization-challenge/data

采集时间：2026-10-19T07:15:25Z

# Data Description

This is a speaker diarization dataset. The goal is to predict **who spoke when** in each audio file. Training data includes audio plus time-aligned speaker labels; test data includes audio only. Submissions list diarization segments for each test file.

## Files

- `train/audio/*.wav` - training audio recordings
- `train/annotation/*.csv` - labels for each training audio (same base filename)
- `test/audio/*.wav` - test audio recordings to predict
- `sample_submission.csv` - example submission in the required format

## Columns

### Training annotations (`train/annotation/*.csv`)

- `start_time` - segment start time in `HH:MM:SS`
- `end_time` - segment end time in `HH:MM:SS`
- `speaker_id` - speaker label (integer, per recording)

### Submission format (`sample_submission.csv`)

- `filename` - test file name 
- `diarization` - JSON list of `{start_time, end_time, speaker_id}` where times are seconds (float) and speaker_id is a label like `SPEAKER_#`

## What you are predicting

For each test audio file, predict a sequence of speaker segments with start/end times and speaker IDs.

## Acronyms

- `SPEAKER_#` - anonymized speaker identifier used in submission output
- `diarization` - the task of segmenting audio by speaker