Dataset · ai4bharat
Labeled speech in 12 Indian languages
ai4bharat/Kathbath
1,684 hours of manually transcribed speech audio across 12 Indian languages for ASR training.
What's inside
The dataset includes audio recordings and text transcripts totaling 1,684 hours (100,000 to 1 million examples). Recordings were gathered from 1,218 speakers across 203 districts of India and cover 12 languages, including Hindi, Bengali, Tamil, Urdu, Sanskrit, and others.
What you can do with it
- Training and fine-tuning automatic speech recognition (ASR) models
- Creating multilingual voice assistants for Indian languages
- Benchmarking speech recognition quality across different dialects
- Size
- 100K–1M rows
- Downloads this month
- 8.8K
- Likes
- 59
- Access
- on request on Hugging Face
Models trained on it
- nvidia/magpie_tts_multilingual_357mText → speech 16.5K
- shunyalabs/zero-stt-hinglishAutomatic speech recognition 706
- udgamlabs/ShrutiAutomatic speech recognition 680
- udgamlabs/shruti-teluguAutomatic speech recognition 403
- onnx-community/zero-stt-hinglish-ONNXAutomatic speech recognition 288
- udgamlabs/shruti-tamilAutomatic speech recognition 277
- udgamlabs/shruti-bengaliAutomatic speech recognition 179
- udgamlabs/shruti-marathiAutomatic speech recognition 165
How to load it
from datasets import load_dataset
ds = load_dataset("ai4bharat/Kathbath")