PineOS.AIProductsSolutionsCase studiesIntegrationsEnterpriseAboutBlogFAQ

← All articles

Voice AI

Voice AI in Telugu, Tamil and Arabic: handling code-switching on real calls

PineOS.AI team · 26 August 2026 · 3 min read

A customer in Hyderabad calls about a loan EMI. She starts in Telugu, says "EMI", "due date" and "bounce charges" in English, reads her account number in a mix of both, and switches to Hindi when the agent misunderstands. That is a normal call, not an edge case. Voice AI built for one clean language at a time breaks on it.

Why real calls are hard

What works in production

  1. Train and test on real call audio. Benchmark on recordings from your own customers, not studio speech. Measure errors separately for numbers and names.
  2. Treat code-switching as the default. Models should expect English terms inside Telugu, Tamil or Arabic speech, rather than forcing one language per call.
  3. Confirm the critical fields. Read back amounts, dates and identifiers before acting on them, in the customer's language.
  4. Let the customer lead the language. Answer in the language the customer is using, and switch when they do.
  5. Hand off cleanly. When confidence drops, pass the call to a person with the transcript and the details already captured.
  6. Keep the data where it belongs. Banks and regulated companies often need speech and transcripts processed in-country.

How PineOS Nexus approaches it

PineOS Nexus runs voice agents for inbound and outbound calls in English, Hindi, Telugu, Tamil and Arabic. Its speech and language models are tuned for Indic languages and Arabic, and are built to handle accents, code-switching and noisy call-centre audio. Models can be served from in-country infrastructure, and Nexus runs in PineOS Cloud (India, UAE or US), your VPC, on-premise or air-gapped.

For a wider comparison of platforms, see our guide to voice AI platforms for banks and NBFCs in India.

Related

Talk to our team