SilencioNetwork/indian-languages-speech

🇮🇳 Indian Languages Speech Dataset

📋 Table of Contents

• Overview
• Off-the-Shelf Availability
• Dataset Statistics
• Language Coverage
• Use Cases
• Dataset Structure
• Data Fields
• Getting Started
• Data Collection
• Quality Assurance
• Limitations & Biases
• License & Usage
• Citation
• Contact

🎯 Overview

The Indian Languages Speech Dataset provides high-quality, real-world speech recordings from native speakers across 9 major Indian languages. We have 16 Indian languages available off-the-shelf, representing over 1.4 billion speakers globally. This dataset is designed for training and evaluating automatic speech recognition (ASR), text-to-speech (TTS), and speaker identification systems for Indic languages.

Key Features

✅ Native speakers only - 100% L1 (first language) speakers
✅ Rich demographic metadata - Gender, age, occupation, location, dialect
✅ Free speech recordings - Spontaneous, natural speech with real-world disfluencies
✅ Real-world acoustic conditions - Natural environments with background noise
✅ Crowdsourced at scale - Collected from 2M+ global contributors
✅ Diverse speaker demographics - Age range 18-70+, balanced gender distribution
✅ Regional dialect coverage - Multiple dialect variants per language

Supported Languages

This dataset includes samples from 9 major Indian languages. We have 16 Indian languages available off-the-shelf (see inventory table below).

Languages with samples in this dataset:

• 🔵 Hindi (hi) - Indo-Aryan, 600M speakers
• 🔵 Bengali (bn) - Indo-Aryan, 265M speakers
• 🔵 Urdu (ur) - Indo-Aryan, 170M speakers
• 🔵 Marathi (mr) - Indo-Aryan, 83M speakers
• 🔵 Telugu (te) - Dravidian, 82M speakers
• 🔵 Gujarati (gu) - Indo-Aryan, 56M speakers
• 🔵 Kannada (kn) - Dravidian, 44M speakers
• 🔵 Odia (or) - Indo-Aryan, 38M speakers
• 🔵 Malayalam (ml) - Dravidian, 38M speakers

Additional Indian languages available OTS (contact for samples): English (India accent), Assamese, Kashmiri, Santali, Nepali, Sanskrit, Sindhi — 7 additional languages with 2,500+ hours available.

🗂️ This is a Sample Dataset

These 276 recordings represent a tiny sample of Silencio's capabilities. This dataset showcases our data quality and collection methodology, but it is not our full inventory.

Full Off-the-Shelf Inventory - All Indian Languages:

Language	Script	OTS Recordings	OTS Hours	In This Sample?
Hindi	Devanagari	184,636	2,000 hours	✅ 41 files
English (India)	Latin	115,609	2,428 hours	❌ Contact
Bengali	Bengali	32,857	319 hours	✅ 45 files
Telugu	Telugu	23,224	293 hours	✅ 25 files
Marathi	Devanagari	14,994	198 hours	✅ 25 files
Urdu	Perso-Arabic	10,395	111 hours	✅ 51 files
Odia	Odia	9,408	105 hours	✅ 21 files
Assamese	Assamese	8,793	65 hours	❌ Contact
Gujarati	Gujarati	7,789	82 hours	✅ 18 files
Kannada	Kannada	7,731	248 hours	✅ 25 files
Malayalam	Malayalam	5,318	87 hours	✅ 25 files
Kashmiri	Perso-Arabic	2,733	35 hours	❌ Contact
Santali	Devanagari/Ol Chiki	1,481	10 hours	❌ Contact
Nepali	Devanagari	736	11 hours	❌ Contact
Sanskrit	Devanagari	55	0.4 hours	❌ Contact
Sindhi	Devanagari/Arabic	37	1 hour	❌ Contact
TOTAL	16 languages	425,756	5,993 hours	276 files (~15 hours)

This sample = 0.06% of available inventory for Indian languages (updated: March 30, 2026)

Samples Available For:

• ✅ Hindi (41 files) - 600M speakers, India's most widely spoken language
• ✅ Bengali (45 files) - 265M speakers, West Bengal & Bangladesh
• ✅ Urdu (51 files) - 170M speakers, Pakistan & India
• ✅ Telugu (25 files) - 82M speakers, Andhra Pradesh & Telangana
• ✅ Marathi (25 files) - 83M speakers, Maharashtra
• ✅ Odia (21 files) - 38M speakers, Odisha
• ✅ Gujarati (18 files) - 56M speakers, Gujarat
• ✅ Kannada (25 files) - 44M speakers, Karnataka
• ✅ Malayalam (25 files) - 38M speakers, Kerala

Available OTS (Contact for Samples):

• English (India accent) - 2,428 hours available
• Assamese - 65 hours
• Kashmiri - 35 hours
• Santali - 10 hours
• Nepali - 11 hours
• Sanskrit - 0.4 hours (ancient/classical language)
• Sindhi - 1 hour

See our other datasets: south-asian-speech for additional South Asian language samples

Silencio's Complete OTS Catalog:

• 📊 3.8M recordings across 1,708 language/country combinations
• ⏱️ 48,000 hours of audio (2,000+ days)
• 🌍 147,000+ unique speakers from 180+ countries
• ✅ Immediate commercial licensing available

On-Demand Sourcing:

• 🌍 Any language, any accent - leveraging our 2M+ contributor network
• 📊 Custom specifications - demographic targeting, acoustic environments, speech types
• ⚡ Rapid turnaround - weeks, not months

Explore more datasets: View our other HuggingFace listings to see the breadth of our capabilities, or contact sofia@silencioai.com for the full off-the-shelf catalog and pricing.

📊 Dataset Statistics

Metric	Value
Total Audio Files	276
Total Languages	9
Total Speakers	200+ unique speakers
Native Speaker Population	1.4B speakers represented
Audio Format	WAV (16-bit PCM)
Sample Rate	16 kHz / 44.1 kHz
Total Duration	~15 hours
Avg. File Duration	5-30 seconds
Geographic Coverage	Pan-India (major language regions)

Language Distribution

Language	Files	Speakers	Primary Regions
Urdu	51	40+	Pan-India, Pakistan
Bengali	45	35+	West Bengal, Bangladesh
Hindi	41	35+	Hindi belt (Delhi, UP, MP, Bihar, etc.)
Telugu	25	20+	Andhra Pradesh, Telangana
Marathi	25	20+	Maharashtra, Goa
Malayalam	25	20+	Kerala, Lakshadweep
Kannada	25	20+	Karnataka
Odia	21	15+	Odisha
Gujarati	18	15+	Gujarat, Mumbai
Total	276	220+	Pan-India

🌍 Language Coverage

Indo-Aryan Languages

Language	ISO 639-1	Speakers	Script	Primary Regions	In Sample?
Hindi	hi	600M	Devanagari	Delhi, UP, MP, Bihar, Rajasthan, Haryana, Uttarakhand, HP, Jharkhand	✅ 41 files
Bengali	bn	265M	Bengali	West Bengal, Bangladesh, Tripura, Assam	✅ 45 files
Urdu	ur	170M	Perso-Arabic	Pan-India, Pakistan	✅ 51 files
Marathi	mr	83M	Devanagari	Maharashtra, Goa, Dadra & Nagar Haveli	✅ 25 files
Gujarati	gu	56M	Gujarati	Gujarat, Daman & Diu, Dadra & Nagar Haveli	✅ 18 files
Odia	or	38M	Odia	Odisha, West Bengal	✅ 21 files
Assamese	as	15M	Assamese	Assam, Arunachal Pradesh	❌ Contact
Kashmiri	ks	7M	Perso-Arabic	Jammu & Kashmir	❌ Contact
Nepali	ne	16M	Devanagari	Sikkim, West Bengal, Nepal	❌ Contact
Sindhi	sd	2.7M	Devanagari/Arabic	Gujarat, Rajasthan, Pakistan	❌ Contact
Sanskrit	sa	-	Devanagari	Classical/liturgical language	❌ Contact

Dravidian Languages

Language	ISO 639-1	Speakers	Script	Primary Regions	In Sample?
Telugu	te	82M	Telugu	Andhra Pradesh, Telangana, Puducherry	✅ 25 files
Kannada	kn	44M	Kannada	Karnataka	✅ 25 files
Malayalam	ml	38M	Malayalam	Kerala, Lakshadweep, Puducherry	✅ 25 files

Other Language Families

Language	ISO 639-1	Speakers	Family	Script	Primary Regions	In Sample?
Santali	sat	7.6M	Austroasiatic	Devanagari/Ol Chiki	Jharkhand, West Bengal, Odisha, Assam	❌ Contact
English (India)	en-IN	125M+	Germanic	Latin	Pan-India (second language)	❌ Contact

Dialect Coverage

The dataset includes multiple regional dialects per language:

• Hindi: Delhi, Mumbai, Patna, Lucknow, Jaipur
• Bengali: Kolkata, Dhaka, Sylhet
• Urdu: Delhi, Hyderabad, Lucknow, Karachi
• Telugu: Hyderabad, Vijayawada, Visakhapatnam
• Marathi: Mumbai, Pune, Nagpur
• Malayalam: Thiruvananthapuram, Kochi, Kozhikode
• Kannada: Bangalore, Mysore, Mangalore
• Odia: Cuttack, Bhubaneswar, Ganjam
• Gujarati: Ahmedabad, Surat, Vadodara

🎯 Use Cases

Automatic Speech Recognition (ASR)

• Train ASR models for 6 major Indian languages
• Evaluate existing models on real-world Indic speech
• Build multilingual ASR systems for South Asian markets
• Develop dialect-aware recognition systems

Text-to-Speech (TTS)

• Train neural TTS models for Indian languages
• Create natural-sounding voice synthesis systems
• Develop multi-speaker TTS with demographic control
• Build dialect-specific TTS engines

Speaker Recognition & Verification

• Train speaker identification models for Indic languages
• Evaluate speaker verification systems on diverse demographics
• Build voice biometrics for multilingual applications

Language Identification

• Train language detection models for South Asian languages
• Distinguish between closely related Indic languages
• Build multilingual language identification systems

Accent & Dialect Research

• Study phonetic variations across Indian regions
• Analyze code-switching patterns in multilingual contexts
• Research sociolinguistic patterns in Indian languages

Model Evaluation & Benchmarking

• Evaluate multilingual models on underrepresented languages
• Benchmark ASR/TTS performance on real-world Indic speech
• Test robustness to background noise and speaker diversity

📁 Dataset Structure

indian-languages-speech/
├── hindi_india/
│   └── free_speech/
│       ├── data/
│       │   ├── recording_*.wav
│       └── metadata.csv
├── bengali_india/
│   └── free_speech/
│       ├── data/
│       └── metadata.csv
├── urdu_india/
│   └── free_speech/
│       ├── data/
│       └── metadata.csv
├── telugu_india/
│   └── free_speech/
│       ├── data/
│       │   ├── audio_*.wav
│       └── metadata.csv
├── marathi_india/
│   └── [same structure as Telugu]
├── malayalam_india/
│   └── [same structure]
├── kannada_india/
│   └── [same structure]
├── odia_india/
│   └── [same structure]
└── gujarati_india/
    └── [same structure]

File Naming Convention

• Audio files: recording_*.wav or audio_*.wav (depending on source)
• Metadata: metadata.csv (one per language, located in {language}_india/free_speech/)

📋 Data Fields

Each audio sample is accompanied by comprehensive metadata in CSV format:

Field	Type	Description
file_name	string	Relative path to audio file
id	integer	Unique recording ID
audio	audio	Audio data (automatically loaded by HF datasets)
gender	string	Speaker gender (male/female/other)
ethnicity	string	Speaker ethnicity
occupation	string	Speaker occupation
country_code	string	ISO country code (IN for India)
birth_place	string	Speaker's birthplace
mother_tongue	string	Speaker's native language
dialect	string	Regional dialect variant
year_of_birth	integer	Birth year (for age calculation)
years_at_birth_place	integer	Years lived at birthplace
languages_data	json	Other languages spoken (with proficiency)
os	string	Recording device OS
device	string	Device type (Mobile/Desktop/Tablet)
browser	string	Browser used for recording
duration	float	Audio duration in seconds
emotions	json	Emotional content labels
language	string	Recording language
location	string	Recording location type (office/outdoor/home)
noise_sources	json	Background noise description
script_id	integer	Script identifier (for prompted speech)
type_of_script	string	Script category (all recordings are free_speech)
script	string	Original script text (Indic script)
transcript	string	Speech transcription (may be "unknown" for free speech)
speaker_id	string	Unique speaker identifier (UUID)

Example Metadata Entry

{
  "file_name": "data/audio_6258613.wav",
  "id": 6258613,
  "gender": "female",
  "ethnicity": "Asian",
  "occupation": "voice over artist",
  "country_code": "IN",
  "birth_place": "India",
  "mother_tongue": "Gujarati",
  "dialect": "Gujarat - Surat",
  "year_of_birth": 1982,
  "years_at_birth_place": 44,
  "languages_data": [{"level": "native", "language": "Gujarati"}],
  "os": "Windows",
  "device": "Desktop",
  "browser": "Chrome",
  "duration": 300.0,
  "emotions": "{happy}",
  "language": "Gujarati",
  "location": "office",
  "noise_sources": "{silence}",
  "script_id": 77581,
  "type_of_script": "free_speech",
  "script": "# તમે ચિત્ર દોરવાનું પસંદ કરો છો?",
  "transcript": "unknown",
  "speaker_id": "e8bf8999-7c74-559c-863d-00a8e5531767"
}

🚀 Getting Started

Installation

pip install datasets soundfile librosa

Basic Usage

from datasets import load_dataset

# Load entire dataset
dataset = load_dataset("SilencioNetwork/indian-languages-speech")

# Load specific language
gujarati = load_dataset("SilencioNetwork/indian-languages-speech", name="gujarati_india")

# Load the free_speech split (all recordings are free speech)
telugu_free = load_dataset(
    "SilencioNetwork/indian-languages-speech",
    name="telugu_india",
    split="free_speech"
)

Iterate Through Examples

from datasets import load_dataset

dataset = load_dataset("SilencioNetwork/indian-languages-speech", name="kannada_india")

for example in dataset["free_speech"]:
    print(f"File: {example['file_name']}")
    print(f"Language: {example['language']}")
    print(f"Speaker: {example['gender']}, {2026 - example['year_of_birth']} years old")
    print(f"Dialect: {example['dialect']}")
    print(f"Duration: {example['duration']}s")
    print(f"Audio shape: {example['audio']['array'].shape}")
    print("---")

Load Audio with Librosa

import librosa
from datasets import load_dataset

dataset = load_dataset("SilencioNetwork/indian-languages-speech", name="malayalam_india")

for example in dataset["free_speech"]:
    # Audio is pre-loaded by HuggingFace datasets
    audio_array = example["audio"]["array"]
    sampling_rate = example["audio"]["sampling_rate"]
    
    # Or load manually with librosa
    audio, sr = librosa.load(example["file_name"], sr=None)
    
    # Process audio
    mfcc = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=13)
    print(f"MFCC shape: {mfcc.shape}")

Filter by Demographics

from datasets import load_dataset

dataset = load_dataset("SilencioNetwork/indian-languages-speech", name="marathi_india")

# Filter female speakers
female_speakers = dataset["free_speech"].filter(lambda x: x["gender"] == "female")

# Filter by age (speakers born after 1990)
young_speakers = dataset["free_speech"].filter(lambda x: x["year_of_birth"] > 1990)

# Filter by location
outdoor_recordings = dataset["free_speech"].filter(lambda x: x["location"] == "outdoor")

print(f"Female speakers: {len(female_speakers)}")
print(f"Young speakers: {len(young_speakers)}")
print(f"Outdoor recordings: {len(outdoor_recordings)}")

Training Example (PyTorch)

import torch
from datasets import load_dataset
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

# Load dataset
dataset = load_dataset("SilencioNetwork/indian-languages-speech", name="telugu_india")

# Load pre-trained model and processor
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base")

# Prepare data
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_values"] = processor(
        audio["array"],
        sampling_rate=audio["sampling_rate"]
    ).input_values[0]
    return batch

dataset = dataset.map(prepare_dataset, remove_columns=["audio"])

# Training loop (simplified)
for batch in dataset["free_speech"]:
    inputs = torch.tensor(batch["input_values"]).unsqueeze(0)
    outputs = model(inputs)
    # ... training logic

🔬 Data Collection Methodology

Collection Platform

All recordings were collected through Silencio's crowdsourcing platform, which enables large-scale, diverse data collection from native speakers globally.

Contributor Requirements

• Native speakers only: All contributors must be L1 (first language) speakers
• Age verification: Contributors 18+ years old
• Device requirements: Smartphone, tablet, or desktop with microphone
• Consent: All contributors provided informed consent for data use

Recording Process

1. Speaker onboarding: Contributors provide demographic information and language proficiency
2. Script assignment: Random assignment of free speech prompts (spontaneous, unscripted speech)
3. Recording: Browser-based or mobile app recording with quality checks
4. Validation: Automated quality checks (duration, amplitude, noise level)
5. Human review: Manual review of flagged recordings

Quality Control Checkpoints

• Minimum duration requirements (5-30 seconds for free speech)
• Amplitude threshold checks to filter silent recordings
• Noise level assessment to ensure audibility
• Language verification by native speaker reviewers
• Duplicate detection and removal

✅ Quality Assurance

Automated Quality Checks

• Audio format validation: All files validated for correct format (WAV, 16-bit PCM)
• Duration checks: Minimum and maximum duration thresholds enforced
• Amplitude analysis: Files with insufficient audio energy flagged and removed
• Metadata completeness: All required fields validated for presence and format

Manual Review Process

• Language verification: Native speakers verify language accuracy
• Transcription spot-checks: Random sample transcriptions reviewed for accuracy
• Demographic validation: Spot-checks on demographic metadata consistency
• Noise assessment: Recordings with excessive noise flagged for review

Known Issues

• Some transcript fields marked as "unknown" for free speech recordings (spontaneous speech)
• Occasional background noise in outdoor and office recordings (intentional for real-world robustness)
• Minor audio clipping in a small subset of recordings (<1%)

⚠️ Limitations & Biases

Geographic Limitations

• Urban bias: Majority of speakers from urban centers (Ahmedabad, Bangalore, Mumbai, Hyderabad, etc.)
• State coverage: Limited coverage from smaller towns and rural areas
• Accent diversity: Regional accents within each state may be underrepresented

Demographic Limitations

• Age distribution: Skewed toward 25-45 age group (typical of crowdsourcing platforms)
• Occupation bias: Overrepresentation of tech workers and voice-over artists
• Education level: Higher education levels than general population (self-selection bias)

Technical Limitations

• Device variability: Mix of mobile and desktop recordings with varying microphone quality
• Sampling rates: Non-uniform sampling rates (16 kHz and 44.1 kHz)
• Background noise: Variable noise levels across recordings (office, outdoor, home)

Language & Linguistic Limitations

• Script coverage: Limited coverage of some script variations and orthographic styles
• Code-switching: Minimal English-Indic code-switching examples
• Dialect representation: Some minor dialects may be underrepresented

Recommendations for Use

• Data augmentation: Apply data augmentation techniques to address demographic imbalances
• Domain-specific fine-tuning: Fine-tune on domain-specific data for production use cases
• Bias mitigation: Monitor model performance across demographic groups and dialects
• Validation on target domain: Test models on your specific use case data before deployment

⚖️ License & Usage

License: CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International)

✅ Allowed Uses

• ✅ Research and academic use
• ✅ Model training for non-commercial applications
• ✅ Educational purposes and teaching
• ✅ Personal projects and experimentation
• ✅ Open-source projects (non-commercial)

❌ Prohibited Uses

• ❌ Commercial products or services
• ❌ Selling or licensing the dataset
• ❌ Commercial ASR/TTS systems without separate licensing
• ❌ Voice cloning for commercial applications
• ❌ Surveillance or law enforcement applications without consent

💼 Commercial Licensing

For commercial licensing inquiries, including:

• Commercial ASR/TTS deployments
• Enterprise voice AI applications
• Voice biometrics systems
• Custom data collection

Contact: sofia@silencioai.com

🏢 About Silencio

Silencio is a voice AI data company providing scaled sourcing of real-world audio and speech data. With 2M+ contributors across 180+ countries, we specialize in:

• ✅ Diverse accent and dialect coverage for 100+ languages
• ✅ Real-world acoustic environments (office, outdoor, home, car)
• ✅ Comprehensive demographic and sociolinguistic metadata
• ✅ On-demand data collection for underrepresented languages
• ✅ Rapid turnaround for custom dataset requirements

Our global contributor network enables us to source:

• Rare languages and dialects
• Specific demographic segments
• Custom recording scenarios
• Large-scale datasets in weeks, not months

Learn more: silencioai.com

📚 Citation

If you use this dataset in your research or applications, please cite:

@dataset{silencio_indian_languages_2026,
  title={Indian Languages Speech Dataset},
  author={Silencio Network},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/SilencioNetwork/indian-languages-speech},
  note={6 major Indian languages (Gujarati, Kannada, Malayalam, Marathi, Odia, Telugu) with native speaker recordings and demographic metadata}
}

Example Citation in Text

"We trained our multilingual ASR model on the Indian Languages Speech Dataset (Silencio Network, 2026), which provides native speaker recordings across 6 major Indian languages with comprehensive demographic metadata."

🤝 Contact

Dataset Issues & Feedback

Found issues with the data? Have suggestions for improvement?

• Email: sofia@silencioai.com
• HuggingFace Community: Discussion Forum

Commercial Inquiries

For commercial licensing, custom data collection, or partnership opportunities:

• Email: sofia@silencioai.com
• Website: silencioai.com

Contributing

We welcome contributions to improve the dataset:

• Report errors: Open a discussion on HuggingFace
• Suggest improvements: Share feedback via email or discussion forum
• Request features: Let us know what additional data or metadata would be valuable

---

Built by Silencio | Licensed under CC BY-NC 4.0

---

Tags: #IndianLanguages #Gujarati #Kannada #Malayalam #Marathi #Odia #Telugu #IndicLanguages #SpeechData #ASR #TTS #Multilingual #VoiceAI #SouthAsia #CrowdsourcedData