The case for Indian-language speech data
Code-mixed, conversational, narrowband: why real calls teach speech models what read speech can't.
Speech models have become remarkably good at the languages that dominate the public internet. Ask them to follow a fast, noisy call between a Bengaluru support agent and a customer who switches between Kannada, Hindi and English inside a single sentence, and the picture changes. The gap isn't a lack of interest. It's a lack of the right data.
Hundreds of millions of speakers, little public audio
India has many languages with tens of millions of speakers each, and a great deal of everyday business happens in them. Yet the speech that is easy to collect at scale (audiobooks, news broadcasts, read-aloud corpora) is a narrow slice of how people actually talk. Read speech is slow, careful and grammatical. Conversation is none of those things.
What labs need is conversational audio: two or more people, overlapping turns, hesitations, interruptions, background noise, and the specific vocabulary of a domain. Contact centres, banks, insurers, hospitals and delivery platforms produce exactly this every day, in every major Indian language, and have often kept years of it.
Code-mixing is the norm, not the exception
A model trained on clean monolingual speech tends to treat code-mixing as an error. In practice it's how many conversations work. An agent might greet a customer in Hindi, read a policy term in English, and confirm an address in Marathi. Numbers, product names and technical terms often stay in English whatever language surrounds them.
Training data has to reflect this, and transcripts have to handle it honestly. That usually means keeping the original script where speakers used it, offering a romanised form alongside, and labelling segments rather than forcing a single language tag on a whole call.
Telephony audio is its own domain
Call-centre audio is narrowband: most of it was recorded at 8 kHz, compressed, and passed through networks that clip and distort. Models that will run on phone lines need to learn from phone lines. Upsampling studio recordings doesn't reproduce the artefacts that matter.
Where the source has them, separate channels for agent and customer are especially useful. They make speaker attribution reliable and let models learn turn-taking without guessing who said what.
Outcomes make conversations teachable
The most useful call data comes with a label for how the call ended: sale or no sale, issue resolved or escalated, a promise to pay or a refusal. Those labels exist because businesses track them. Paired with the audio, they let a lab train and evaluate models on what actually happened, not just on what was said.
Privacy has to be designed in
Conversational audio is full of personal information: names, phone numbers, addresses, card and account numbers, Aadhaar and PAN numbers, policy and claim IDs. None of it should leave the owner's systems. That means transcribing locally, detecting those spans, and replacing them with tone or silence in the audio and tags in the text, all before export, with a sample reviewed and signed off by the owner.
India's Digital Personal Data Protection framework sets out duties for anyone processing personal data, and owners need confidence that licensing doesn't put them on the wrong side of it. De-identifying at the source is the most direct way to keep personal data where it already is.
What good looks like
A strong Indian-language speech dataset is conversational, domain-specific, and honest about its mix of languages. It keeps the telephony characteristics models will meet in deployment, carries outcome labels where the business recorded them, and arrives de-identified with a clear account of how. Most of that already exists inside organisations that run calls at scale. The work is in licensing it carefully, so that it can finally be used.