Voice bots for people who do not read: language is the easy part
Speech models now handle many Indian languages well. What decides whether a voice assistant works is everything around the language.
Amogh Suman
For a large share of Indian customers of financial services, an app is the wrong interface. They may not read comfortably, may not read the language the app is written in, or may use a basic phone. Voice looks like the answer, and with speech recognition and translation in Indian languages improving fast, it is tempting to think the problem is solved.
It is not. Language has become the easy part. The hard parts are trust, structure and what happens when the conversation goes wrong.
Trust
A borrower who receives a call about their loan has learned, often painfully, that calls about money can be scams. A voice assistant has to prove it is real before it asks for anything. That means a known number, a voice and script the customer recognises, and a firm rule that the assistant never asks for a PIN or an OTP.
The irony is that a well-designed bot can be more trustworthy than a human caller, because it follows the rule every time. It has to be designed for that from the first word.
Structure
Reading lets people skim, go back and compare. Listening does not. A voice assistant has to work with how memory works:
- one question at a time
- short choices, three at most
- numbers read slowly and repeated, especially amounts and dates
- a summary before any commitment
A loan statement that fits on one screen can take minutes to read aloud. Deciding what to leave out is a design task, and it is harder than translation.
A voice assistant has to be designed for how people listen, not translated from how they read.
Mixed language and real conditions
Customers switch languages mid-sentence, use local words for financial terms, and call from noisy streets. Demos usually take place in a quiet room with a clear speaker. Testing should take place where customers actually are.
The way out
Every voice assistant needs a quick, obvious way to reach a person. For many customers that is the most important feature. A bot that handles routine questions well and hands over the rest cleanly is better than one that tries to handle everything and traps people in a loop.
Voice can bring many more people into formal finance. It will do that only if it is built around the customer's situation, not around the capabilities of the latest speech model.