Voice is how a very large number of Indians prefer to interact with a service, particularly outside the metros, and it is where most product teams' assumptions break most completely. A speech system that performs well in a demonstration meets accented speech, background noise, code-switching mid-sentence and a call quality that varies by the second, and its accuracy falls off a cliff that no benchmark predicted.
The first thing to accept is code-mixing, because it is the norm rather than an edge case. Indian speakers routinely move between English and a regional language within one sentence, using English for numbers, technical terms and brand names inside a Hindi or Tamil or Bengali sentence structure. A recogniser configured for a single language will transcribe half the utterance and guess at the rest. Choosing a model that handles the mixture, and testing specifically on mixed speech, matters more than the headline accuracy figure for either language alone.
Design the flow to be tolerant rather than precise. Open-ended prompts — how can I help you today — produce responses a recogniser will struggle with, and users who do not know what they are permitted to say. Constrained prompts with a small number of expected answers, confirmation of anything consequential, and an easy path to a human, produce systems that work at the accuracy you actually have rather than the accuracy you wished for. Always offer keypad entry alongside speech for numbers, because digits are where recognition errors are most costly and most common.
Numbers deserve specific attention. Amounts, account numbers, dates and PIN codes are exactly what people call about, and exactly what a speech system gets wrong — particularly with Indian number conventions, where a caller may say lakhs or use a mixture of English and vernacular digits. Confirm anything numeric by reading it back, and accept keypad correction without restarting the flow.
Measure the right thing, which is task completion rather than recognition accuracy. A system with ninety per cent word accuracy that fails to complete the caller's task is worse than one with eighty per cent that gets them to the right outcome, because the caller does not care what was transcribed. Track how many calls reach a resolution, how many escalate to a human, and where in the flow people abandon — the abandonment point is a design problem, not a model problem, and it is fixable without touching the speech stack.
One more thing worth stating plainly: an IVR that traps people is worse than no IVR. Callers in India are frequently calling because something has gone wrong and they are already frustrated, and a system that will not let them reach a person converts a support problem into a churn problem. Offer the human path early, make it obvious, and treat the proportion of callers who take it as information about your automated flow rather than as a failure of the caller.
