BORENTIS

Conversation intelligence

Code-Switched Speech Recognition for Retail Floors: Hinglish, Regional Languages and the Showroom Floor

An retail conversation is rarely in one language. The advisor opens in Hindi, the customer answers in Hinglish, the model name is English, the price is in lakhs, the objection is in Marathi or Tamil or Bengali, and all of it is switched mid-sentence over showroom music. English-first speech recognition, which is what most conversation intelligence tools were built on, turns that into noise, and every score built on the transcript is fiction. This page explains what code-switching does to transcription, why it matters more in retail than in a contact centre, and the five things to ask a vendor before trusting any number built on their transcript.

What code-switching does to transcription

  • Language identification fails inside a sentence: 'EMI kitna banega, processing fee ke saath?' is Hindi grammar around English nouns, and a model that picks one language per utterance loses half of it.
  • Named entities are English inside Hindi: model names, scheme names, variants, competitors. Those are the words a sales score depends on, and the ones a Hindi-only model mangles.
  • Numbers come in two systems: lakhs and crores, and thousands; 'paanch lakh' and 'five lakh' and '5L' are one price.
  • Regional languages are not dialects of Hindi: a Marathi, Tamil, Telugu, Kannada, Bengali or Gujarati floor needs its own model, switching with English and often with Hindi.
  • Noise: music, three conversations, a generator; the phone is a metre from two speakers, not on a headset.

Why it matters more in retail than in a contact centre

A contact-centre call is one language, one channel, a headset, and a script. A showroom conversation is the opposite on every count, and the scoring is more demanding: a sales score needs to know that the finance offer was quoted correctly, which needs the number and the scheme name right, not just the gist. A transcript that is mostly right is enough for sentiment; it is not enough for offer accuracy or a disclosure record. The category this serves is defined at /learn/what-is-retail-conversation-intelligence/.

Five things to ask a vendor

Borentis transcribes Hinglish, Hindi, English and regional languages, code-switched, with the engine abstracted behind one interface so that a client can route to an approved provider; the detail is at /technology/. The capture that produces the audio is at /learn/how-to-capture-sales-interactions-at-retail-stores/.

  1. Transcribe a recording from my own busiest store, at peak hour, in my floor's languages. Not a demo clip.
  2. Show me the model names, the scheme name and the price in that transcript, correct.
  3. Which regional languages have you transcribed on a real floor, and can I hear one?
  4. What happens to the score when the transcript is uncertain: is it flagged, or scored anyway?
  5. Can the transcription engine be swapped per client, so that a brand with an approved provider can use it?

Frequently asked questions

What is code-switched speech recognition?

Transcribing speech that switches between languages inside a sentence, such as Hindi grammar around English product names, with a regional language mixed in. It is the normal condition of an retail conversation and the case most speech recognition was not built for.

Why does it matter for conversation intelligence?

Because every score depends on the transcript. A model that mangles the model name, the scheme or the price cannot tell whether the offer was quoted correctly, and a sales score built on it is fiction.

How do I test a vendor?

Give them a recording from your own busiest store at peak hour and read the transcript for the model names, the scheme and the price. If those are right the rest usually is.

Related reading

Where Borentis applies this

Borentis is the Agentic Operating System for Customer Interactions, built for Indian retail floors: consented one-tap capture on the advisor's phone, every conversation scored against your playbook with the evidence behind every number, leads created when a number is heard, and coaching from your own best conversations.