Speech data for 10 Indian languages
Spontaneous two-person conversation in Hindi, Bengali, Urdu, Punjabi, Gujarati, Marathi, Tamil, Telugu, Kannada and Malayalam: 8,500 hours from 850 native speakers at $65 to $75 per hour of audio. Bengali was recorded in Bangladesh, Urdu in Pakistan and the other 8 in India.
Compare the Indian language datasets
| Language | Recorded in | Hours | Speakers | Per hour |
|---|---|---|---|---|
| Hindi | India | 2,000 | 200 | $65 |
| Bengali | Bangladesh | 2,000 | 200 | $65 |
| Urdu | Pakistan | 1,000 | 100 | $65 |
| Punjabi | India | 500 | 50 | $65 |
| Gujarati | India | 500 | 50 | $65 |
| Marathi | India | 500 | 50 | $65 |
| Tamil | India | 500 | 50 | $75 |
| Telugu | India | 500 | 50 | $65 |
| Kannada | India | 500 | 50 | $65 |
| Malayalam | India | 500 | 50 | $65 |
| All 10 | 3 countries | 8,500 | 850 | $65 to $75 |
List prices in US dollars per hour of audio. Need another language, more hours or your own recording setup? We collect speech data to spec.
What differs between these languages
Hindi
Everyday Hindi is full of English, with speakers switching mid-sentence (Hinglish), so a model has to handle two languages at once.
Bengali
Bengali has many aspirated consonants and nasal vowels, and Bangladeshi Bengali differs from Kolkata Bengali in pronunciation and vocabulary.
Urdu
Urdu is written in Perso-Arabic script, usually without short vowels, so one written word can be read several ways.
Punjabi
Punjabi is tonal, which is rare for an Indo-European language: tone changes the meaning of words that would otherwise sound the same.
Gujarati
Gujarati has breathy vowels that change meaning and are hard for models trained on other languages to hear.
Marathi
Marathi has sounds like the retroflex "ळ" and, in many words, pronounces "च" and "ज" as "ts" and "dz", which Hindi-trained models get wrong.
Tamil
Spoken Tamil is very different from written Tamil, so a model trained on formal text or read speech struggles with real conversation.
Telugu
Telugu builds long words by adding suffixes and has many retroflex and aspirated consonants. Speech varies across Telangana, coastal Andhra and Rayalaseema, and conversational Telugu borrows English heavily.
Kannada
Kannada builds long words from suffixes, and the spoken language differs noticeably from the written one.
Malayalam
Malayalam joins words together in writing, which gives it some of the longest words of any Indian language.
What every dataset includes
| Speech | Spontaneous two-person conversations recorded in-country, not read scripts |
|---|---|
| Audio | Dual-channel, one channel per speaker |
| Transcripts | Time-aligned and human-verified |
| Metadata | Age, gender, region and recording device for every speaker |
| Consent | Every speaker signed explicit commercial-use consent; the full consent chain is provided for legal review |
| License | Commercial use, including training and evaluating ASR, TTS and other voice AI models |
| Price | $65 to $75 per hour of audio |
Built for
Training and evaluating speech models for India and South Asia:
- Automatic speech recognition (ASR / STT)
- Text-to-speech & voice synthesis prosody
- Voice agents and IVR
- Speaker diarization & turn-taking models
- Speech LLM / audio-understanding fine-tuning
Frequently asked questions
Where were the Bengali and Urdu datasets recorded?
The Bengali dataset was recorded in Bangladesh and the Urdu dataset in Pakistan; the other 8 languages were recorded in India. Bangladeshi Bengali differs from Kolkata Bengali in pronunciation and vocabulary, and Pakistani speakers mix Urdu with English and Punjabi. If you need Bengali or Urdu speakers in India, we record them to spec.
Can I license several Indian languages together?
Yes. Each dataset is licensed at its own published price, with volume and multi-language discounts on request. All 10 share one format: dual-channel audio of spontaneous two-person conversation, time-aligned and human-verified transcripts, and the same speaker metadata, so they combine into one multilingual training set.
How much does Indian language speech data cost?
List prices are per hour of audio: $75 for Tamil; $65 for the other 9. Each dataset page shows its price, and volume and multi-language discounts are available on request.
Can you record other Indian languages?
Yes. Other languages, such as Odia, Assamese or Konkani, can be recorded to spec by native speakers, scripted or spontaneous, with time-aligned transcripts and documented consent. We confirm feasibility and send a quote within one business day.
Hear the languages before you buy
Tell us which languages you need and share your company email. We’ll reach out shortly with representative audio, matching transcripts and the metadata schema, so you can check quality before you commit.