Why Is OpenAI Betting on Voice for ChatGPT's Next Billion Users?
OpenAI's Vijay Raji explains why voice, not text, is the interface that will bring ChatGPT to its next billion users, especially in India.

Why does OpenAI think voice matters more than text for future growth?
ChatGPT has reached 1.2 billion users through an interface that’s almost comically simple: a text box and a button. That worked because typing in English, or in a language with a keyboard that matches your alphabet, is a low-friction way to talk to a model. But according to Vijay Raji, OpenAI’s CTO of Applications, that same interface becomes a barrier for the next wave of users. Most of the next billion people won’t be most comfortable typing in English, and for many of them, typing at all in their native language is harder than just talking. Voice removes that friction entirely.
TL;DR
- OpenAI’s growth team is betting on voice as the default modality for ChatGPT’s next billion users, not text, because typing in English or in Indic scripts is a bigger barrier than most Western users realize.
- Raji’s observations from a recent trip to India showed people defaulting to voice messages even in apps like WhatsApp, rather than typing, which he sees as a signal for how ChatGPT needs to evolve.
- Script and keyboard complexity for Indic languages makes typing meaningfully harder than speaking, which pushes voice from a “nice to have” to a core requirement for adoption in those markets.
- Raji prototyped a voice-first ChatGPT experience himself using Codex, opening the microphone immediately on app launch instead of waiting for a backend connection, before bringing the idea to his product team.
- The shift reflects a broader pattern inside OpenAI where leaders increasingly build working prototypes with AI coding tools rather than pitching ideas verbally, compressing the gap between idea and validation.
- This growth push is happening alongside internal engineering strain, as AI-assisted coding has pushed so much code through OpenAI’s systems that the company had to rebuild its code repository infrastructure to keep up.
What’s actually broken with ChatGPT’s current interface for global growth?
The text box isn’t broken for the audience ChatGPT already has. It’s broken for everyone else. Raji frames this bluntly: English isn’t going to be the most comfortable language for the next billion users, and typing isn’t going to be their most comfortable modality. That’s a different problem than translation. OpenAI can and does localize ChatGPT into other languages, but localization doesn’t solve the deeper issue of input method. Typing in Hindi, Tamil, Bengali, or other Indic languages on a mobile keyboard is slower and more error-prone than typing in English on a QWERTY layout designed for it. Raji calls this out directly as a “script issue.”
The evidence he points to isn’t abstract. He describes watching how people in India already communicate on platforms like WhatsApp: when given the choice, they send voice messages to friends instead of typing texts. That’s a strong signal about preferred modality that has nothing to do with ChatGPT specifically. It’s a pattern baked into how people already use their phones. If that’s how people choose to communicate with each other, a product that only accepts typed input is fighting against an established habit rather than working with it.
How is OpenAI approaching this inside the product?
Rather than simply assigning the idea to a product team, Raji built a working prototype himself using Codex, OpenAI’s coding agent. The prototype stripped ChatGPT down to a simple voice-first experience: when the app opens, it immediately activates the microphone so the user can start speaking while the connection is still being established in the background, then plays back the recorded conversation once the connection completes. The goal was to remove the dead time most voice interfaces have between “open app” and “start talking,” since that gap is exactly where users default back to typing or give up.
Raji is clear that this is a prototype, not a shipped product, and that turning it into something reliable at scale is “a completely different set of problems.” But the exercise mattered because it let him validate the idea for himself before taking it to his team, compressing what used to be a purely verbal pitch into something people could actually try.
This is part of a broader shift Raji describes in how OpenAI’s leaders work. Where a leader used to connect dots across teams and hand off ideas for others to explore, AI coding tools now let that same leader build a rough version of the idea directly. Code, in his words, has become “so accessible” that it’s no longer the bottleneck between having an idea and showing whether it works.
Is voice really the right bet, or just an easy demo?
It’s worth being skeptical about any single leader’s prototype shaping company direction, but the reasoning behind the bet holds up on its own terms. Voice interfaces solve a specific, well-documented friction point: input method mismatch for non-Latin scripts and lower-literacy or lower-typing-speed populations. This isn’t a novel insight unique to ChatGPT. It’s the same logic that has driven voice features in messaging apps, voice search on mobile, and voice assistants generally in markets where typing is a weaker second-choice behavior compared to speaking.
What makes this particular bet notable is where it sits in OpenAI’s growth strategy. Raji’s team is explicitly responsible for pushing ChatGPT’s user base past its current 1.2 billion, and his framing suggests OpenAI sees voice not as a feature addition but as a potential default experience, something closer to the primary way new users encounter the product rather than an alternative mode buried in settings. That’s a meaningfully different design decision than just adding a microphone button to the existing text interface.
The open question is execution. A prototype that opens the mic instantly is a good user-experience idea, but delivering accurate, low-latency voice understanding across dozens of Indic languages and dialects, at the reliability users expect from a product used by over a billion people, is a much harder engineering problem than the demo suggests.
Why does this growth push coincide with OpenAI’s internal engineering strain?
The voice bet doesn’t exist in isolation. Raji also described a separate but related problem: OpenAI’s internal engineering systems have been pushed to their limits by AI-assisted coding. Token usage tied to coding tools has grown 17 times over the past year at frontier AI labs, according to Raji, compared to roughly 2x at typical software companies. That surge in output created downstream bottlenecks in code review, CI/CD, and testing, and eventually broke the assumptions behind OpenAI’s code repository system itself. Git, built for developers who work offline and periodically push changes, wasn’t designed for a world where AI agents are checking out and pushing code constantly. OpenAI built a new internal system from scratch to handle that throughput.
Why does this matter for the voice story? Because it shows the same dynamic playing out at two different scales inside OpenAI: once ideas become cheap to prototype and ship, the bottleneck moves somewhere else. For engineering, the bottleneck moved from “writing code” to “reviewing and merging it safely.” For growth, the bottleneck is moving from “can we build a voice feature” to “can we validate, localize, and scale it across hundreds of millions of new users with different languages and expectations.”
Frequently Asked Questions
Why does OpenAI think voice is more important than text for growth in India specifically?
Raji’s reasoning is that typing in English or in Indic scripts is a higher-friction task than speaking for many users, and he points to existing behavior, like people already preferring voice messages over typed texts on WhatsApp, as evidence of that preference.
Did OpenAI actually build a voice-first version of ChatGPT?
Raji built a prototype himself using Codex, OpenAI’s coding agent, that opens the microphone immediately when the app launches rather than waiting for a backend connection. He describes it as a working prototype used to validate the idea internally, not a shipped product.
How many users does ChatGPT currently have?
According to Raji, ChatGPT has reached 1.2 billion users as of this conversation, using what he describes as a simple text box and button interface.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Is this voice push connected to OpenAI’s other engineering challenges?
Indirectly. OpenAI has separately had to rebuild its internal code infrastructure because AI-assisted coding increased engineering output faster than its existing systems (including Git-based workflows) could handle. Both situations reflect the same pattern: once a process gets cheap or easy, the bottleneck shifts elsewhere.
Does voice replace typing in ChatGPT, or add to it?
The transcript doesn’t specify final product plans, only that Raji’s team is treating voice as a candidate for the default experience for new users, particularly those for whom English typing or Indic-script typing is a barrier.
