RailKosh: Indian Railway PA Announcement Benchmark

RailKosh is a bilingual benchmark dataset of Indian railway public address (PA) announcements for ASR evaluation and named entity recognition research. Audio is embedded directly in parquet.

Quick Start

from datasets import load_dataset

ds    = load_dataset('raianand/RailKosh')            # both splits
ds_en = load_dataset('raianand/RailKosh', split='en')
ds_hi = load_dataset('raianand/RailKosh', split='hi')

# Streaming
ds = load_dataset('raianand/RailKosh', streaming=True, split='hi')

Dataset Summary

Split	Language	Clips	Audio	Station names	Announcement type
en	English	1,000	TTS-synthesized	English	English
hi	Hindi	513	Field recordings	Hindi	Hindi
Total		1,513

Announcement types — EN: Arrival · Departure · Standing · Delay · Cancelled · Platform Change

Schema

Field	Type	Description
clip_id	string	Unique identifier (e.g. clip_1)
language	string	en or hi
audio	Audio	Embedded MP3, decoded at 22050 Hz
transcript	string	Gold transcript in the clip's language
train_no	string	5-digit zero-padded train number
train_name	string	Train name in the clip's language
from_station	string	Origin station in the clip's language
to_station	string	Destination station in the clip's language
via_stations	string	Intermediate stations (comma-separated)
announcement_type	string	Announcement type in the clip's language
platform	string	Platform number

Demo

🚂 RailVaani Live Demo

Citation

@inproceedings{railkosh2025,
  title     = {RailKosh: A Benchmark for Hindi Railway Announcement ASR},
  author    = {Rai, Anand Kumar and others},
  booktitle = {Proceedings of EMNLP 2025 (Demo Track)},
  year      = {2025}
}