szzs1693/coswara-data

Coswara COVID-19 Audio Dataset

This is the derived dataset of Coswara Project Dataset, with audio normalized by downsampled to 16 kHz for storage efficiency. The dataset has been converted to Parquet format, and hosted on Hugging Face for more accessibility.

Dataset Description

The Coswara dataset is a crowdsourced collection of respiratory sounds and speech recordings for COVID-19 diagnosis research, collected by the Indian Institute of Science (IISc) Bangalore. The dataset contains audio recordings of breathing, cough, and speech from 2,746 participants collected between April 2020 and February 2022.

Dataset Summary

• Total Participants: 2,746
• Total Audio Files: 24,718 (9 recordings per participant)
• Collection Period: April 2020 - February 2022
• Audio Format: 16kHz WAV (resampled from mixed source rates: 48kHz, 44.1kHz, 16kHz)
• Geographic Coverage: 20+ countries (91.6% India)
• COVID-19 Positive: 681 participants (24.8%)

Supported Tasks

• Audio Classification: COVID-19 detection from respiratory sounds
• Multi-task Learning: Joint prediction using multiple audio types
• Health Status Classification: 8-class health status prediction
• Symptom Analysis: Correlation between audio features and COVID symptoms

Dataset Structure

Configurations

The dataset provides two configurations (subsets):

1. metadata Configuration

Participant health and demographic information.

Splits:

• train: 1,922 participants
• val: 412 participants
• test: 412 participants

Features: 36 fields including COVID status, demographics, symptoms, and comorbidities

2. audio Configuration

Individual audio recordings with quality annotations and denormalized metadata.

Splits:

• train: ~17,298 audio files (1,922 participants × 9)
• val: ~3,708 audio files (412 participants × 9)
• test: ~3,708 audio files (412 participants × 9)

Features: Audio arrays (16kHz) + quality scores + 10 key metadata fields

Audio Types

Each participant provided 9 recordings:

1. breathing-deep - Deep breathing sounds
2. breathing-shallow - Shallow breathing sounds
3. cough-heavy - Heavy cough sounds
4. cough-shallow - Shallow cough sounds
5. vowel-a - Vowel 'a' pronunciation
6. vowel-e - Vowel 'e' pronunciation
7. vowel-o - Vowel 'o' pronunciation
8. counting-normal - Counting at normal speed
9. counting-fast - Counting at fast speed

Schema

metadata Configuration

Field	Type	Description
id	string	Unique participant identifier (hash)
a	int64	Age in years
g	string	Gender (male/female/other)
covid_status	string	Health status (8 categories, see below)
record_date	date	Recording submission date
test_status	string	COVID test result (p=positive, n=negative, na=not applicable, ut=under test)
test_date	date	Date of COVID test
testType	string	Test type (RAT/RT-PCR)
vacc	string	Vaccination status (y=fully, p=partially, n=not vaccinated)
l_c	string	Country
l_s	string	State/province
l_l	string	Locality/city
ep	string	English proficiency (y/n)
rU	string	Returning user (y/n)
Symptoms
fever	string	Fever symptom
cough	string	Cough symptom
bd	string	Breathing difficulties
loss_of_smell	string	Loss of smell symptom
st	string	Sore throat
ftg	string	Fatigue
mp	string	Muscle pain
diarrhoea	string	Diarrhea symptom
Comorbidities
smoker	string	Smoking status
cold	string	Cold
ht	string	Hypertension
diabetes	string	Diabetes
asthma	string	Asthma
ihd	string	Ischemic heart disease
cld	string	Chronic lung disease
pneumonia	string	Pneumonia
others_resp	string	Other respiratory conditions
others_preexist	string	Other pre-existing conditions
CT Scan
ctScan	string	CT scan performed
ctDate	date	CT scan date
ctScore	string	CT severity score
Quality Scores
breathing_deep_quality	int64	Quality label (0=bad, 1=good, 2=excellent)
breathing_shallow_quality	int64	Quality label
cough_heavy_quality	int64	Quality label
cough_shallow_quality	int64	Quality label
vowel_a_quality	int64	Quality label
vowel_e_quality	int64	Quality label
vowel_o_quality	int64	Quality label
counting_normal_quality	int64	Quality label
counting_fast_quality	int64	Quality label

audio Configuration

Field	Type	Description
participant_id	string	Unique participant identifier (joins with metadata)
audio_type	string	Type of recording (breathing-deep, cough-heavy, etc.)
audio	Audio	Audio array at 16kHz sampling rate
quality_score	int64	Manual quality annotation (0=bad, 1=good, 2=excellent)
record_date	date	Recording submission date
covid_status	string	Participant's COVID health status
age	int64	Participant's age
gender	string	Participant's gender
test_status	string	COVID test result
vacc	string	Vaccination status
country	string	Participant's country

COVID Status Categories

Status	Description	Count	Percentage
healthy	Healthy individuals	1,433	52.2%
positive_mild	COVID-19 positive with mild symptoms	426	15.5%
no_resp_illness_exposed	Exposed but no respiratory illness	248	9.0%
positive_moderate	COVID-19 positive with moderate symptoms	165	6.0%
resp_illness_not_identified	Respiratory illness, cause unknown	157	5.7%
recovered_full	Fully recovered from COVID-19	146	5.3%
positive_asymp	COVID-19 positive, asymptomatic	90	3.3%
under_validation	Status under validation	81	3.0%

Quality Score Distribution

Quality labels were manually annotated for each audio file:

• 0 (Bad): Poor quality, significant noise/distortion (excluded from classification)
• 1 (Good): Acceptable quality for analysis
• 2 (Excellent): High quality, clear audio

Example distribution (breathing-deep):

• Quality 2: 71.5%
• Quality 1: 12.1%
• Quality 0: 16.5%

Usage Examples

Load Metadata Only

from datasets import load_dataset

# Load metadata configuration
metadata = load_dataset("szzs1693/coswara-data", "metadata")

# Access splits
train_metadata = metadata["train"]
val_metadata = metadata["val"]
test_metadata = metadata["test"]

# Convert to pandas for analysis
import pandas as pd
df = train_metadata.to_pandas()

# Analyze COVID status distribution
print(df['covid_status'].value_counts())

Load Audio for COVID Classification

from datasets import load_dataset

# Load audio configuration
audio_ds = load_dataset("szzs1693/coswara-data", "audio")

# Filter for high-quality cough recordings
cough_ds = audio_ds["train"].filter(
    lambda x: x['audio_type'] in ['cough-heavy', 'cough-shallow']
              and x['quality_score'] >= 1
)

# Separate by COVID status
covid_positive = cough_ds.filter(
    lambda x: x['covid_status'].startswith('positive')
)
healthy = cough_ds.filter(
    lambda x: x['covid_status'] == 'healthy'
)

print(f"COVID-19 positive samples: {len(covid_positive)}")
print(f"Healthy samples: {len(healthy)}")

Train Audio Classifier with Wav2Vec2

from datasets import load_dataset
from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2ForSequenceClassification

# Load breathing recordings
ds = load_dataset("szzs1693/coswara-data", "audio")
breathing_ds = ds["train"].filter(
    lambda x: x['audio_type'] == 'breathing-deep' and x['quality_score'] >= 1
)

# Create binary labels (COVID positive vs healthy)
def create_labels(example):
    example['label'] = 1 if example['covid_status'].startswith('positive') else 0
    return example

breathing_ds = breathing_ds.map(create_labels)

# Load pretrained model
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-base")
model = Wav2Vec2ForSequenceClassification.from_pretrained(
    "facebook/wav2vec2-base",
    num_labels=2
)

# Preprocess audio
def preprocess(batch):
    audio = [x["array"] for x in batch["audio"]]
    inputs = feature_extractor(
        audio,
        sampling_rate=16000,
        padding=True,
        return_tensors="pt"
    )
    return inputs

# ... continue with training loop

Multi-Task Learning (All Audio Types)

from datasets import load_dataset

# Load all audio types
audio_ds = load_dataset("szzs1693/coswara-data", "audio")

# Filter for high quality
train_ds = audio_ds["train"].filter(lambda x: x['quality_score'] >= 1)

# Group by participant for multi-task learning
from collections import defaultdict

participant_data = defaultdict(list)
for example in train_ds:
    participant_data[example['participant_id']].append(example)

# Each participant now has up to 9 audio samples
print(f"Participants with complete recordings: {sum(1 for v in participant_data.values() if len(v) == 9)}")

Join Audio with Full Metadata

from datasets import load_dataset
import pandas as pd

# Load both configurations
metadata = load_dataset("szzs1693/coswara-data", "metadata")
audio = load_dataset("szzs1693/coswara-data", "audio")

# Convert to pandas
metadata_df = metadata["train"].to_pandas()
audio_df = audio["train"].to_pandas()

# Join for full metadata access
merged = audio_df.merge(
    metadata_df,
    left_on='participant_id',
    right_on='id',
    how='left',
    suffixes=('', '_full')
)

# Now you have all 36 metadata fields for each audio sample
print(f"Merged dataset shape: {merged.shape}")

Dataset Splits

Splits are stratified by covid_status to maintain class distribution:

• Train (70%): 1,922 participants, ~17,298 audio files
• Val (15%): 412 participants, ~3,708 audio files
• Test (15%): 412 participants, ~3,708 audio files

Important: All 9 audio recordings from the same participant are kept in the same split to prevent data leakage.

Data Collection

Collection Method

• Platform: Web-based submission
• Period: April 13, 2020 - February 24, 2022
• Participants: Crowdsourced volunteers (within India and to a smaller extend from outside India)
• Recording Environment: Uncontrolled (home/personal devices)
• Consent: All participants provided informed consent

Quality Control

• Manual quality annotation by trained annotators
• 3-point scale: 0 (bad), 1 (good), 2 (excellent)
• Annotations available for filtering low-quality samples

Known Limitations

1. Uncontrolled Recording Environment: Varying background noise, device quality
2. Mixed Sample Rates: Original recordings have variable sample rates (48kHz, 44.1kHz, 16kHz) due to different recording devices; all resampled to 16kHz for consistency
3. Class Imbalance: 52% healthy vs 25% COVID-positive
4. Geographic Bias: 91.6% from India
5. Self-Reported Data: Some metadata fields rely on participant reporting
6. Temporal Coverage: Primarily pre-vaccination era (2020) and Delta variant period (2021)
7. Missing Data: Some audio files may be NULL due to collection errors (~1-2%)

Known Data Quality Issues

• Null Audio: 392/24,714 recordings (1.6%) are null due to:

  • 2 files missing from source data
  • 390 files corrupt/unreadable by librosa

• Missing Metadata: Many participants have incomplete metadata:

  • test_status: 51.5% missing
  • vacc: 64.9% missing

Users should filter null values when using the dataset:

# Filter out null audio
ds_clean = ds.filter(lambda x: x['audio'] is not None)

# Filter out missing test status
ds_tested = ds.filter(lambda x: x['test_status'] is not None)

Citation

If you use this dataset, please cite:

Original Paper

@inproceedings{Sharma_2020, series={interspeech_2020},
   title={Coswara — A Database of Breathing, Cough, and Voice Sounds for COVID-19 Diagnosis},
   url={http://dx.doi.org/10.21437/Interspeech.2020-2768},
   DOI={10.21437/interspeech.2020-2768},
   booktitle={Interspeech 2020},
   publisher={ISCA},
   author={Sharma, Neeraj and Krishnan, Prashant and Kumar, Rohit and Ramoji, Shreyas and Chetupalli, Srikanth Raj and R., Nirmala and Ghosh, Prasanta Kumar and Ganapathy, Sriram},
   year={2020},
   month=oct, collection={interspeech_2020} }

Nature Scientific Data Publication

@article{bhattacharya2023coswara,
  title={Coswara: A respiratory sounds and symptoms dataset for remote screening of SARS-CoV-2 infection},
  author={Bhattacharya, Debarpan and Sharma, Neeraj Kumar and Dutta, Debottam and Chetupalli, Srikanth Raj and Mote, Pravin and Ganapathy, Sriram and Chandrakiran, C and Nori, Sahiti and Suhail, KK and Gonuguntla, Sadhana and Alagesan, Murali},
  url={https://doi.org/10.1038/s41597-023-02266-0}
  DOI={10.1038/s41597-023-02266-0}
  journal={Scientific data},
  volume={10},
  number={1},
  pages={397},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

Links

• Paper (arXiv): https://arxiv.org/abs/2005.10548
• Paper (Nature): https://www.nature.com/articles/s41597-023-02266-0
• Project Website: https://coswara.iisc.ac.in/

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.

Ethical Considerations

• All participants provided informed consent for data collection and research use
• Participant identifiers are anonymized hash strings
• Geographic data is limited to country/state/city level
• Researchers should consider geographic and demographic biases when generalizing findings
• This dataset is for research purposes only and should not be used for clinical diagnosis without proper validation

Acknowledgments

This dataset was collected and curated by the Indian Institute of Science (IISc) Bangalore. We thank all the volunteers who contributed their recordings to support COVID-19 research.