README.md

---
pretty_name: Snow Mountain
language:
- hi
- bgc
- kfs
- dgo
- bhd
- gbk
- xnr
- kfx
- mjl
- kfo
- bfz
annotations_creators:
- 'null': null
language_creators:
- 'null': null
multilinguality:
- multilingual
source_datasets:
- Snow Mountain
task_categories:
- automatic-speech-recognition
- text-to-speech
task_ids: []
configs:
- hi
- bgc
dataset_info:
- config_name: hi
  features:
  - name: Unnamed
    dtype: int64
  - name: sentence
    dtype: string
  - name: path
    dtype: string
  splits:
  - name: train_500
    num_examples: 400
  - name: val_500
    num_examples: 100
  - name: train_1000
    num_examples: 800
  - name: val_1000
    num_examples: 200
  - name: test_common
    num_examples: 500
  dataset_size: 71.41 hrs
- config_name: bgc
  features:
  - name: Unnamed
    dtype: int64
  - name: sentence
    dtype: string
  - name: path
    dtype: string
  splits:
  - name: train_500
    num_examples: 400
  - name: val_500
    num_examples: 100
  - name: train_1000
    num_examples: 800
  - name: val_1000
    num_examples: 200
  - name: test_common
    num_examples: 500
  dataset_size: 27.41 hrs
license: cc-by-sa-4.0
---

# Snow Mountain

## Dataset Description

- **Paper: https://arxiv.org/abs/2206.01205**
- **Point of Contact: Joel Mathew**

### Dataset Summary

The Snow Mountain dataset contains the audio recordings (in .mp3 format) and the corresponding text of The Bible (contains both Old Testament (OT) and New Testament (NT)) in 11 Indian languages. The recordings were done in a studio setting by native speakers. Each language has a single speaker in the dataset. Most of these languages are geographically concentrated in the Northern part of India around the state of Himachal Pradesh. Being related to Hindi they all use the Devanagari script for transcription. 

We have used this dataset for experiments in ASR tasks. But these could be used for other applications in speech domain, like speaker recognition, language identification or even as unlabelled corpus for pre-training.

### Supported Tasks and Leaderboards

Atomatic speech recognition, Speech-to-Text, Speaker recognition, Language identification

### Languages

Hindi, Haryanvi, Bilaspuri, Dogri, Bhadrawahi, Gaddi, Kangri, Kulvi, Mandeali, Kulvi Outer Seraji, Pahari Mahasui, Malayalam, Kannada, Tamil, Telugu 

## Dataset Structure
```
data
  |- cleaned
    |- lang1
      |- book1_verse_audios.tar.gz
      |- book2_verse_audios.tar.gz
        ...
        ...
      |- all_verses.tar.gz
      |- short_verses.tar.gz
    |- lang2
      ...
      ...   
  |- experiments 
    |- lang1
      |- train_500.csv
      |- val_500.csv
      |- test_common.csv
        ...
        ...
    |- lang2
      ...
      ...
  |- raw
    |- lang1
      |- chapter1_audio.mp3
      |- chapter2_audio.mp3
        ...
        ...
      |- text
        |- book1.csv
        |- book1.usfm
          ...
          ...
    |- lang2
      ...
      ...
```
### Data Instances

A data point comprises of the path to the audio file, called `path` and its transcription, called `sentence`. 

```
{'sentence': 'क्यूँके तू अपणी बात्तां कै कारण बेकसूर अर अपणी बात्तां ए कै कारण कसूरवार ठहराया जावैगा',
 'audio': {'path': 'data/cleaned/haryanvi/MAT/MAT_012_037.wav',
  'array': array([0., 0., 0., ..., 0., 0., 0.]),
  'sampling_rate': 16000},
 'path': 'data/cleaned/haryanvi/MAT/MAT_012_037.wav'}
```

### Data Fields

`path`: The path to the audio file

`audio`: A dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: `dataset[0]["audio"]` the audio file is automatically decoded and resampled to `dataset.features["audio"].sampling_rate`. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. `dataset[0]["audio"]` should always be preferred over `dataset["audio"][0]`.

`sentence`: The transcription of the audio file.

### Data Splits

We create splits of the cleaned data for training and analysing the performance of ASR models. The splits are available in the `experiments` directory.  The file names indicate the experiment and the split category. Additionally two CSV files are included in the data splits - `all_verses` and `short_verses`. Various data splits were generated from these main two CSVs. `short_verses.csv` contains audios of length < 10s and corresponding transcriptions. `all_verses.csv` contains complete cleaned verses including long and short audios. Due to the large size (>10MB), we keep these CSVs compressed in the `tar.gz format in the `cleaned` folder.

## Dataset Loading
`raw` folder has chapter wise audios in .mp3 format. For doing experiments, we might need audios in .wav format. Verse wise audio files are keept in the `cleaned` folder in .wav format. This results in a much larger size which contributes to longer loading time into memory. Here is the approximate time needed for loading the Dataset.
  - Hindi (OT books): ~20 minutes
  - Hindi minority languages (NT books): ~9 minutes
  - Dravidian languages (OT+NT books): ~30 minutes

## Details
Please refer to the paper for more details on the creation and the rationale for the splits we created in the dataset.

### Licensing Information

The data is licensed under the Creative Commons Attribution-ShareAlike 4.0 International Public License (CC BY-SA 4.0)


### Citation Information

Please cite this work if you make use of it:

```
@inproceedings{Raju2022SnowMD,
  title={Snow Mountain: Dataset of Audio Recordings of The Bible in Low Resource Languages},
  author={Kavitha Raju and V. Anjaly and R. Allen Lish and Joel Mathew},
  year={2022}
}
```