Accent conversion is a speech AI model that changes the accent of a speaker's voice while they speak, and keeps their words, their tone and much of their own voice. It is used mainly in call centres. An agent in Hyderabad or Manila who is speaking with customers in the United States is heard with an accent closer to the one the caller is used to.
India handles a large share of the world's customer-service calls. In call centres in Hyderabad, Bengaluru, Gurugram and other cities, agents answer calls on behalf of banks, telecom companies and retailers in the United States. Some callers there find certain Indian accents harder to follow on a phone line, so calls take longer and need more repeats. Call centres have long sent their agents to accent training. Accent conversion attempts the same result with software.
How does it work? The software sits between the agent's headset and the phone line. It receives the agent's voice in short pieces, about 20 to 40 milliseconds long. A speech recognition step works out which speech sounds were said. A second model then produces the same sounds again, with the pronunciation of the target accent, in a voice close to the agent's own. The caller hears the result about a fifth of a second later. The words are not changed, and nothing is translated. The caller's own voice reaches the agent unchanged.
Imagine an agent in a Hyderabad call centre who handles billing calls for an electricity company in the United States. A customer in Ohio calls about a bill of $186. The agent says, 'Your bill was raised on the thirtieth.' The software processes her speech in 30-millisecond pieces. It changes the sounds that American callers most often mishear, such as those in 'thirtieth', and keeps the rest. The customer hears her about 200 milliseconds after she speaks. He does not ask her to repeat anything, and the call ends sooner.
Accent conversion raises fair questions. Some people say it protects agents from callers' prejudice. Others say it hides who the agent is. And because the technology changes voices, it is close to voice cloning, which is why deepfake detection tools must be able to tell such voices apart. It is often used together with a predictive dialler on outgoing calls. Of all the speech and voice AI products, it has the most direct link to India's services industry.