Home / Data types / Speech
Real call-centre speech data for speech AI
Datadhara sources real customer support and sales call recordings, with transcripts checked by people, speaker turns and intent labels. Phone numbers, names and account details are masked, and every call is licensed for AI training with consent from the source.
What we can source
| Record type | Format | Typical labels |
|---|---|---|
| Support calls | WAV / FLAC + JSON transcript | Speaker turns, timestamps, intent |
| Sales calls | WAV / FLAC + JSON transcript | Speaker turns, outcome |
| Voice notes | Audio + transcript | Language, topic |
We collect to your spec, so volumes, languages and labels are agreed per project. Tell us what you need.
Example record
{
"speaker": "customer",
"t": "00:14.2",
"text": "order [REDACTED] still not delivered, can you check?",
"lang": "code-mixed",
"intent": "delivery_delay"
}
Values are illustrative. Real deliveries follow the schema agreed in your spec.
How personal information is removed
Names, phone numbers, order and account numbers are masked in both the transcript and the audio. Calls that mention sensitive personal details beyond what the spec needs are dropped. See the full methodology.
What teams use it for
- Speech recognition (ASR) for accented, noisy and code-mixed speech
- Intent detection and call summarisation
- Evaluation sets for voice agents
How licensing works
Every source signs a written agreement that covers AI training. You start with a small pilot and a dataset card, and if the data works for you we agree volume, timeline and license terms (exclusive or non-exclusive) in writing.
Questions
Which languages are available?
It depends on the sources for your project. We work with low-resource and code-mixed speech as well as English. Tell us the languages you need and we'll confirm before scoping.
Are transcripts machine-generated?
Transcripts are drafted automatically and then checked and corrected by people. The dataset card states the review process.
Tell us what your models are missing
A few lines is enough: the domain, the format, and roughly how much.
Request data →