38,000 hours of transcribed two-speaker conversations in 22 Indian languages from consented contributors, with dialect, age band, environment and device metadata.
Data product · Listing updated 11 Nov 2026
India
Transcript word error rate below 4% on audited samples
38,000 hours, about 1,200 new hours a month
Startups, Mid-market, Enterprises, Academia and research
Contracted contributors with signed recording releases covering ASR, TTS and language-model training
Mobile-app conversations on guided topics, 16 kHz WAV, two-pass human transcription with code-switch tagging
Natural, unscripted conversations recorded for speech recognition and synthesis training, with documented consent for every voice.
Contributors are recruited across 180 districts, paid per session and sign a recording release that covers model training. Recordings are 16 kHz WAV captured on mobile phones in homes, streets, vehicles and offices. Each file is transcribed by one linguist and reviewed by a second; audited samples show a transcript word error rate below 4%. Names, phone numbers and addresses spoken in conversation are bleeped and tagged. The data is licensed for ASR, TTS and language-model training only. It may not be used for speaker identification, voice cloning of contributors or any biometric purpose. About 1,200 new hours are added each month.
14 fields as documented by Kathana Language Data.
| Field | Type | Description | Example |
|---|---|---|---|
| utterance_id | string | Identifier of the time-aligned segment | — |
| language | string | ISO 639-3 language code | — |
| dialect_region | string | Dialect or region of the speaker | — |
| speaker_id | string | Pseudonymous speaker identifier | — |
| gender | string | Self-reported gender | — |
| age_band | string | Age band such as 25-34 | — |
| audio_duration_sec | number | Length of the segment in seconds | — |
| sample_rate_hz | integer | Audio sample rate (16,000) | — |
| recording_environment | string | quiet_home, street, vehicle or office | — |
| device_type | string | Phone class used for recording | — |
| transcript_native | string | Verbatim transcript in native script | — |
| transcript_romanized | string | Romanised transcript | — |
| code_switch_tags | json | Spans where speakers switch language | — |
| consent_record_id | string | Reference to the signed recording release | — |
Real rows from the provider's sample file. Personal data is masked before anything is shown.
| utterance_idstring | languagestring | dialect_regionstring | genderstring | age_bandstring | audio_duration_secstring | transcript_romanizedstring |
|---|---|---|---|---|---|---|
| utt_000184 | hin | Awadhi belt | female | 25-34 | 6.4 | kal market mein sabzi bahut mehngi thi |
| utt_000185 | tam | Kongu | male | 35-44 | 4.9 | naan naalaikku office-ku late-aa varuven |
| utt_000186 | ben | Rarhi | female | 18-24 | 5.2 | ami ekhon bus stop-e dariye achhi |
| utt_000187 | mar | Varhadi | male | 45-54 | 3.8 | paus ala tar shet kaam thambel |
Prices in USD, as published by Kathana Language Data
per transcribed audio hour, minimum 100 hours
5,000-hour multi-language bundle, perpetual licence
Converted amounts use DataStoa display rates and are approximate. Kathana Language Data quotes the final price, taxes and terms.
Request detailed pricing1 country.
Time coverage: 2018 to present.
Provider-reported
Latency: Transcript word error rate below 4% on audited samples
DataStoa has not audited these figures. Ask the provider how they are measured, or test them on a sample.
The licence is agreed directly between you and Kathana Language Data. This summary is provider-reported; the linked terms prevail.
Not listed: internal analytics, AI retrieval and grounding, redistribution allowed, marketing and outreach. Ask Kathana Language Data if you need any of these.
Licensed Indic speech and text data for AI, from consented contributors
Kathana Language Data builds speech and text datasets for Indian languages with a network of more than 40,000 consented contributors and in-house linguists in Pune and Hyderabad. AI teams license its corpora and commission annotation, RLHF and custom collection projects for ASR, TTS and language models.
Also available on: AWS Data Exchange, Databricks Marketplace
What buyers say about Kathana Language Data. Reviews cover the provider, not only this listing.
from 2 reviews
The 2,000-item preference pilot came back on time with agreement metrics and annotator notes in Hindi and English. Guideline changes were absorbed within a day. We have since scaled to several thousand comparisons a week.
Meera J., Applied scientist, Enterprise AI lab
Every hour of audio links to a consent record that covers commercial model training, and the speaker metadata let us balance dialects properly. Transcription quality in Tamil and Bengali was the best we have bought.
Vikram R., Speech ML lead, Voice assistant company