Picture a phone call with your bank. You start in English, slip into Hindi halfway through a sentence, and then say your name — a name no voice assistant has ever pronounced quite right. Most voice bots stumble right there. They pause, they mishear, they ask you to repeat yourself, and the conversation that should have taken two minutes stretches into ten. Bengaluru-based Arrowhead AI has built a voice to fix exactly that moment.
Illustration of a customer on a live voice-AI call, the kind of conversation Arrowhead AI’s Fletch is built for.
A voice built for how India actually speaks
Arrowhead AI launched Fletch on 30 October 2026, its own in-house text-to-speech model for multilingual customer conversations across India and Southeast Asia. The company says Fletch speaks 14 languages, including English, Hindi, Tamil, Telugu and Bengali, plus Malay and Indonesian, and it can switch between languages in the middle of a sentence — the way millions of Indians actually talk.
That switching is the heart of the engineering. A customer discussing an EMI payment might move between Hindi and English inside a single breath, drop in a product name, and then say a place name from their hometown. The company says Fletch is designed to handle regional accents, code-mixed sentences and Indian proper nouns without losing the thread, and to adjust its tone to the conversation — with 12 emotion styles ranging from calm and empathetic to firm and apologetic, so a collections reminder does not sound like a sales pitch and a support call does not sound like a robot reading a script.
Fast enough to feel like a person
Speed is where a voice model either feels human or feels like a machine. The company says Fletch runs at a reported latency of 80 milliseconds and begins generating speech in under 100 milliseconds — fast enough that there is no awkward silence while the caller waits for a reply. Founder and CEO Devyani Gupta puts it simply: “Fletch is our first step towards voice AI that sounds like it belongs here: 14 languages, and it starts speaking in under 100 milliseconds. We want every customer to hear a conversation, not a machine.”
Underneath that speed is a DiTAR-based architecture — a diffusion-transformer approach to speech generation — adapted and trained by Arrowhead itself. The company says the model was trained on thousands of hours of curated, use-case-specific regional speech, alongside recordings from professional voice artists. Co-founder and CTO Vengadanathan Srinivasan says, “We built on the latest research in speech generation and trained it on thousands of hours of regional speech, with real voice artists. The result runs at 80 milliseconds of latency and adjusts its tone to what’s being said.”
More Read
Already live on a million calls a day
Fletch is not a demo waiting for its first customer. The company says the model is already deployed across live customer calls, powering more than one million calls a day for clients including Axis Bank, RBL Bank, IndusInd Bank, Tata 1mg, upGrad, KreditBee, Kissht and TVS Credit. It handles collections, sales, renewals and inbound customer service, and the company says live deployments keep refining how it pronounces the terminology, accents and speaking patterns found in real conversations.
The company says Fletch delivers around 30% better pronunciation accuracy on selected names and terms — brand names, product names, customer names and Indian proper nouns — which means fewer moments where a customer has to spell their name twice or ask the voice to repeat itself. For a bank or a lender handling lakhs of calls, that small improvement, multiplied across every conversation, is the difference between a customer who hangs up frustrated and one who finishes the call feeling heard.
Owning every layer of the conversation
Fletch also completes a bigger bet Arrowhead has been building since it was founded in 2022. The company says it now runs its own speech-recognition system, its own text-to-speech model and its own small language model, plus a dedicated model for addresses — so every layer of the conversation is built for how India speaks, rather than assembled from third-party speech tools tuned for somewhere else. Owning the voice layer, the company says, lets it tune every millisecond and every emotion for its markets, instead of accepting whatever a generic model happens to sound like.
That matters far beyond banking. The same voice that can take a loan-renewal call in Tamil, switch to English for a product name and back again can carry the same warmth into healthcare, education and consumer services across India and Southeast Asia. Arrowhead is now extending Fletch further into the region, with Malay and Indonesian support, and plans to demonstrate the model at the Singapore FinTech Festival 2026, scheduled for 18–20 November at Singapore EXPO.
India has spent years teaching machines to understand its languages. With Fletch, Arrowhead AI is teaching them to speak those languages back — quickly, warmly, and in the customer’s own mix of words. The next time a voice on the phone gets your name right on the first try, in the language you actually speak, there is a good chance it will be a voice built in Bengaluru.
Keep reading: The Proof Behind Every ‘Yes’: Inside SecureDApp’s SecureCMS, India’s Consent Platform for Banks · From Hostel Room to Self-Driving Lab: Inside NovAtom’s AI for Scientific Discovery · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
