When Sharath Keshava Narayana landed his first job, he was asked to don the American identity of Nathan.
As an 18-year-old call centre executive, Narayana was required to speak in an American accent efficiently so that the customer on the other end of the call did not realise he was speaking to an Indian.
When BPOs started in India two decades back, callers from India routinely donned American identities and spoke in fake accents to customers, who had no clue that the person they were speaking to, was sitting thousands of miles away on a different continent.
"When I joined the call centre, the first thing I was asked was to go through an accent training course. I was told that you cannot go on a call and say that my name is Sharad. I was told to introduce myself as Nathan," Narayana said, reminiscing his job at a call centre he took as a side gig as an engineering student.
However, Narayana did not pursue the call-centre job after he was asked to change his name to Nathan.
Years later, after co-founding another speech startup Observe and working with the online medical platform Practo, Narayana is now the co-founder of Silicon Valley-based startup Sanas, which converts Asian accents into American or British.
How does it work?
Demonstrating the capabilities of his venture to Moneycontrol, Narayana called up an Indian colleague.
“So Vikram, why don’t you enable the Sanas accent convertor?” asked Narayana.
“I know you used to work in a call centre before, so why don’t you give us some background on what you were working on.”
As they were chatting, Vikram said, “I have just enabled Sanas, and right now I am speaking through Sanas.”
As he spoke, Vikram’s voice changed into that of an American.
Apart from the first few seconds into the colleague's sentences where the algorithm seemed to be struggling to have an effect, for the remainder of the around two-minute conversation, the American accent was on point and convincing.
Over the last few days, his product that "transforms accents of Asians into that of White Americans" is being accused of being ‘racist’.
But Narayana is unfazed. He says the criticism and the controversy were expected and insists that his product is helping customer-care service agents in protecting themselves against discrimination and bias.
"Well, racism is a small part of it. The major chunk of the western population has a hard time understanding South American or South Asian accents," Narayana said.
Narayana cited the example of a client who currently uses Sanas for elderly care. “They’re actually the sweetest people, they’re very patient and they love to talk,” he said.
“But the challenge is sometimes they don’t understand what customer agents are saying. They keep asking: Can you repeat? Because they don’t understand the accent.”
“So I thought if we can build a product that can make people understand very easily and seamlessly, it will just make the whole customer service experience better,” he said.
Sriram Rajagopal, founder and director at Diamondpick, a provider of talent solutions to customers in the digital technology and operations industry, concurred.
“The whole issue here is around comprehension – that is not just the way you speak, but also how you understand. The challenge is that a caller may not follow what is being spoken and as a result will be unable to respond quickly.
“So even if you had a translator or a converter, it needs to be real-time. So if it's real-time, then it gets easier. Then you don't start figuring out where the person is from, you can just have a conversation,” Rajagopal said.
However, Rajagopal said that rather than converting speech into a specific accent, it would be easier to make it accent neutral.
Sanas began in 2020 as a project of Stanford graduates – Venezuelan national Andrés Pérez Soderi, Chinese national Shawn Zhang and Russian national Maxim Serebryakov, who are currently the co-founders of Sanas.
The reason behind establishing the startup is also similar to that of Narayana’s experience in the call centre.
“Raul H. Garcia Letona is one of my closest friends, and a fellow Stanford University student,” Serebryakov said in a LinkedIn post. “We started Sanas for him. During school, he was forced to take a leave of absence and return home to Nicaragua.”
Narayana recalled this story during his interaction with Moneycontrol.
“He started looking for jobs with high English proficiency. But he only had his high school diploma. With those credentials, the highest paying job he could find was in a call centre,” Serebryakov wrote.
“Working there, he faced daily bias associated with his accent. He was regularly verbally abused and discriminated against over the phone,” he said, adding that Raul was forced to leave his job after that.
Based on Raul’s experiences, the idea behind Sanas first materialised as a research paper written by the three for Stanford. While they were working on their research, the trio were also reaching out to investors for funds.
Narayana’s first engagement with the startup was as an angel investor.
He closely followed the developments of the startup and progress on its products, and finally, in 2021, Narayana took the decision to effectively play a much larger role in the company and joined as a co-founder.
“So I joined them as the fourth co-founder. And because they were just still building it as a research project, I helped them build it like a company.”
“We hired a team and raised capital. And then we picked a few early call centres, because I knew a lot of these call centres enterprises very well. So I helped them do some POC (proof of concept),” he said.
Narayana had experience in this field through his work in Observe, a startup which analysed the speech of customer service agents and gave recommendations on how to improve and so on.
“We went to India, Philippines. And I was personally motivated to do this in India because in the early 2000s a lot of customer service jobs went to the Philippines. For me it was an opportunity to bring these jobs back to India with the help of Sanas,” Narayana said.
How does it work?
Since its establishment, Sanas, over a year, has spent $7 million on building the product. The startup’s artificial intelligence does real-time speech synthesis at the phonetic level of a word.
The AI does not hear the whole word but hears "a bunch of sound", as Narayana puts it. Then the AI maps how that particular sound will be pronounced in say the United Kingdom or United States of America or Philippines, etc.
Initially, India was not among the plans for Sanas. “Everybody told us that building a product for India will be the hardest thing to do,” Narayana said.
When compared to a country like the Philippines where there are only three to four major languages with about 15 dialects, India has hundreds of languages.
“And in our research, we found that India has close to 500 different dialects. So building a machine learning model for such a large dataset is very complex,” he said.
For generating the data that was used to train the AI algorithm, Sanas made use of studios in India, the US, the UK, Latin America and the Philippines, where people were invited to talk to the software.
“So, if someone said “hi, how are you doing today”? in a studio in India, we would have somebody from London say the same thing with the same intonation,” Narayana said, adding that the algorithm has been trained with 10 million hours of audio generated from 40 different studios across the world.
As of now, Narayana said that they have developed Phase 1 of the product, which is a many-to-one accent converter, where the accent input can be numerous and the customer service agent can choose what the output dialect should sound like.
“I always knew that this would be controversial. There is a reason why we were working in stealth mode,” he said.
Before the third week of August, Sanas’ internet presence was only that of a website that currently hosts a three-month-old demo of how Sanas works and a few press releases.
In the demo, a customer service agent and a customer talk to each other in an Indian accent, and the accent quickly changes into American when Sanas is enabled.
The Sanas-enabled voice on the website is robotic and the difference between the demonstration Narayana provided to Moneycontrol and the demo on the website is marked.
Narayana stressed that the product is still not complete. “We were working under the radar because we are working on making the algorithm work two ways.”
“By the end of the year, the developed product will automatically convert the dialect based on what the input dialect is. For instance, if there is a call centre employee in India and he is listening to an American customer, he can change the customer’s accent to Indian,” he said.
Now that Sanas is all over the internet, it has invited criticism by many Western media outlets including SFGate, which termed Sanas a “buzzy Bay Area startup that wants to make the world sound whiter”.
Narayana also does not mince his words when it comes to defending the accusation of Sanas' tech being 'racist'.
“It's easy for people to sit outside and say: We should accept people the way they are. We would love it if that happens. But if you look at an agent's life, he takes 70 to 80 calls in a day. And about at least 15-20 times, somebody will comment on his race, accent, him being brown or him being coloured.”
“It's not a happy feeling for an agent. And that is why if you see in a call centre, people quit in 30-60 days constantly because they just can't take that constant discrimination,” he added.
Now with Sanas’ product live in Indian call centres in cities such as Gurgaon, Bengaluru, Mangalore, Mysore, Gandhinagar, Ahmedabad and so on, Narayana claims that he has been getting favourable feedback from call centre executives.
“Agents have come and told me that for the last 30 days they have not been abused. This makes me happy, and makes me do my job better,” he said.
B2B for now
Sanas is currently available as a B2B app and has been deployed in call centres belonging to enterprises in several cities in India. They have also gone live in the Philippines.
“In the app, all they have to do is change their microphone settings from default microphone to Sanas’ microphone.”
“And after they enable it, they won’t even know that the software is there. They just talk normally, but the person on the other side will understand them more clearly,” he added.
However, Sanas is wary of making the app available for general citizens, or as a B2C app.
There have been numerous cases in India where scammers posing as customer agents with foreign accents, in cities such as Kolkata, Gurgaon and other places, have defrauded foreigners.
“We had a lot of demand from Google and Apple Play stores to release it as a B2C app. But we saw that this could be a problem. So until we know how to do a KYC of a person, we decided it would not be the right thing to release it as a B2C app,” he said.
“That's why we chose to do a B2B, where we only work with large enterprises, like it could either be at a very large call centre, or it could be a large enterprise.”
“And the licence cannot be used apart from the infrastructure within which can be used. We don't want to cause harm to anybody,” he added.
In March 2022, Sanas raised $32 million from a clutch of investors including Insight Venture Partners and Google Ventures.
With the funding, Narayana wants to expand the application for internal communications of enterprises that have teams in different parts of the world.
Last week, the startup launched an office in Bengaluru with 12 employees. Overall, the company has 43 people, and it aims to grow to over 100 by the end of the year, with 50 employees in India.
In the next six months, Sanas will also expand into Latin America, South Africa and the Western Asian region.In the future, Narayana, who was born and brought up in Karnataka, hopes that Sanas will be a reason for bringing back a lot of customer service jobs that shifted to the Philippines in the early 2000s to India as well as in establishing customer centres in Tier-2 or 3 cities such as Kanpur, Bhopal where people will not have to worry about their accent, and Sanas’ technology can help do their job.