In contrast to textual content, which is comparatively uniform, spoken language is richly-layered—with cultural nuances, colloquialisms and emotion. Startups constructing voice-first AI fashions are actually doubling down on one factor above all else: the depth and variety of datasets.
Why voice is rising because the frontline interface
In India, the place oral custom performs a pivotal function in communication, voice isn’t only a comfort—it’s a necessity. “We’re not an English-first or perhaps a text-first nation. Even after we sort in Hindi, we regularly use the English script as an alternative of Devanagari. That’s precisely why we have to construct voice-first fashions—as a result of oral custom performs such a significant function in our tradition,” mentioned Abhishek Upperwal, chief government officer (CEO) of Soket AI Labs.
Voice can also be proving crucial for customer support and accessibility. “Voice performs an important function in bridging accessibility gaps, notably for customers with disabilities,” mentioned Mahesh Makhija, chief, expertise consulting, at EY.
“Many shoppers even want voicing complaints over typing, just because speaking feels extra direct and human. Furthermore, voice is much extra frictionless than navigating cellular apps or interfaces—particularly for customers who’re digitally-illiterate, older, or not fluent in English,” mentioned Makhija, including that “speaking in vernacular languages opens entry to the subsequent half a billion shoppers, which is a serious focus for enterprises.”
Startups like Gnani.ai are already deploying voice methods throughout banking and monetary companies to streamline buyer assist, help with mortgage functions, and get rid of digital queues. “One of the best ways to achieve individuals—no matter literacy ranges or demographics—is thru voice within the native language, so it is crucial to seize the tonality of the conversations,” mentioned Ganesh Gopalan, CEO of Gnani.ai.
The hunt for wealthy, real-world information
As of mid-2025, India’s AI panorama exhibits a transparent tilt towards text-based AI, with over 90 Indian corporations energetic within the area, in comparison with 57 in voice-based AI. Textual content-based platforms are inclined to concentrate on doc processing, chat interfaces, and analytics. In distinction, voice-based corporations are extra concentrated in customer support, telephony, and regional language entry, based on information from Tracxn.
When it comes to funding, voice-first AI startups have attracted bigger funding rounds at later phases, whereas textual content AI startups present broader distribution, particularly at earlier phases.
For instance, Skit.ai, a voice-first AI agency, raised a complete of $47.6 million throughout 5 funding rounds. Equally, Yellow.ai has cumulatively secured round $102 million, together with a serious $78.15M Sequence C spherical in 2021, making it one of many top-funded startups in voice AI, information from Tracxn exhibits.
Nevertheless, information stays the foundational problem for voice fashions. Voice AI methods want large, numerous datasets that not solely cowl completely different languages, but in addition regional accents, slangs and emotional tonality.
Chaitanya C., co-founder and chief technological officer of Ozonetel Communications, put it merely: “The datasets matter probably the most—talking as an AI engineer, I can say it is not about the rest; it is all in regards to the information.”
IndiaAI Mission has allotted ₹199.55 crore for datasets—nearly 2% of the mission’s complete ₹10,300 crore finances —whereas 44% has gone to compute. “Investments solely in compute are inherently transient—their worth fades as soon as consumed. Alternatively, investments in datasets construct sturdy, reusable property that proceed to ship worth over time,” mentioned Chaitanya.
He additionally emphasised the shortage of wealthy, culturally-relevant information in regional languages like Telugu and Kannada. “The quantity of information simply out there in English, when put next with Telugu and Kannada or Hindi, it’s not even comparable,” he mentioned. “Someplace it is simply not excellent, it wouldn’t be pretty much as good as an English story, which is why I wouldn’t need it to inform a Telugu story for my child.”
“Some film comes out, no one’s going to write down it in authorities paperwork, however individuals are going to speak about it, and that’s misplaced,” he added, mentioning that authorities datasets usually lack cultural nuance and on a regular basis language.
Gopalan of Gnani.ai agreed. “The colloquial language is commonly very completely different from the written type. Language consultants have an important profession path forward of them as a result of they not solely perceive the language technically, but in addition know the best way to converse naturally and grasp colloquial nuances.”
Startups are actually using artistic strategies to fill these gaps. “First, we accumulate information instantly from the sphere utilizing a number of strategies—and we’re cautious with how we deal with that information. Second, we use artificial information in some circumstances. Third, we increase that artificial information additional. As well as, we additionally leverage a considerable quantity of open-source information out there from universities and different sources,” Gopalan mentioned.
Artificial information is artificially-generated information that mimics real-world information to be used in coaching, testing, or validating fashions.
Upperwal added that Soket AI makes use of the same strategy: “We begin by coaching smaller AI fashions with the restricted actual voice information we’ve. As soon as these smaller fashions are moderately correct, we use them to generate artificial voice information—primarily creating new, synthetic examples of speech.”
Nevertheless, some intend to consciously steer clear of artificial information.
Ankush Sabarwal, CEO and founding father of CoRover AI, mentioned the corporate depends solely on actual information, intentionally avoiding artificial information, “If I’m a client and I’m interacting with an AI bot, the AI bot will grow to be clever by the advantage of it interacting with a human like me.”
The moral labyrinth of voice AI
As corporations start to scale their information pipelines, the brand new Digital Private Knowledge Safety (DPDP) Act will form how they accumulate and use voice information.
“The DPDP legislation emphasizes three key areas: it mandates clear, particular, and knowledgeable consent earlier than amassing information. Second, it enforces objective limitation—information can solely be used for authentic, said functions like KYC or employment, not unrelated mannequin coaching. Third, it requires information localization, which means crucial private information should reside on servers in India,” mentioned Makhija.
He added, “Corporations have begun together with consent notices in the beginning of buyer calls, usually mentioning AI coaching. Nevertheless, the precise means of how this information flows into mannequin coaching pipelines continues to be evolving and can grow to be clearer as DPDP guidelines are absolutely applied.”
Outsourcing voice information assortment raises purple flags, too. “For a deep-tech firm like ours, voice information is among the strongest types of IP (mental property) we’ve, and outsourcing it might compromise its integrity and possession. What if somebody is utilizing copyrighted materials?” mentioned Gopalan.
========================
AI, IT SOLUTIONS TECHTOKAI.NET