Arrowhead’s Fletch Targets the Friction in Multilingual Voice AI

Voice AI has a pronunciation problem.

Arrowhead AI launched Fletch, an in-house text-to-speech model built for multilingual customer calls across India and Southeast Asia. The Bengaluru-based voice AI startup says Fletch powers more than one million calls daily, with support for 14 languages, in-sentence language switching and speech generation that begins in under 100 milliseconds.

That combination targets the parts of voice automation customers notice first: whether an agent says names correctly, responds without awkward pauses and sounds suited to the conversation. A machine that misunderstands a regional name or turns an apology into a sales pitch has not solved customer service. It has only automated the irritation.

Built for regional speech, not generic demos

Fletch is designed to handle regional accents, code-mixed sentences and Indian proper nouns. Arrowhead says the model supports languages including English, Hindi, Tamil, Telugu, Bengali, Malay and Indonesian, while switching between languages within a sentence.

The company claims approximately 30% better pronunciation accuracy on selected terms, including brand names, product names, customer names and Indian proper nouns. Arrowhead said that improvement can reduce cases in which customers ask voice agents to repeat themselves—a small interaction that can expose a weak system faster than any product demonstration.

Fletch has been trained on thousands of hours of curated regional speech and recordings from professional voice artists. That regional training is intended to improve performance in real-world applications, where callers rarely speak in the neat, carefully paced sentences used in polished demos.

The model uses a DiTAR-based architecture adapted and trained by Arrowhead. It operates alongside the company’s speech recognition system, language model and dedicated address model within Arrowhead’s voice AI platform, connecting speech generation with the other systems that manage a customer conversation.

Latency, tone and live deployment

Arrowhead reports 80 milliseconds of latency, while Fletch begins generating speech in under 100 milliseconds. The model is deployed across live customer calls involving collections, sales, renewals and inbound customer service, giving the performance claims a working environment rather than a laboratory backdrop.

Fletch also adapts its vocal tone to the context of a conversation and incorporates 12 emotional styles, including empathetic, apologetic, calm and firm. The goal is not to make every call theatrical; it is to keep the system’s delivery aligned with what the customer is hearing and saying.

“Owning our voice layer means we can tune every millisecond and every emotion for how India and Southeast Asia actually speak,” Devyani Gupta, Founder and CEO of Arrowhead AI, said.

Gupta described Fletch as the company’s first step toward voice AI that sounds local, highlighting its 14 languages and speech generation in under 100 milliseconds. “We want every customer to hear a conversation, not a machine,” she said.

Vengadanathan Srinivasan, Co-founder and CTO of Arrowhead AI, said the team trained Fletch 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,” he said.

Arrowhead works with enterprises including Axis Bank, RBL Bank, IndusInd Bank, Tata 1mg, upGrad, KreditBee, Kissht and TVS Credit. Those relationships place Fletch in customer interactions where pronunciation, timing and tone affect collections, sales, renewals and service outcomes rather than serving as decorative features on a specification sheet.

The company is expanding Fletch’s capabilities across Southeast Asia, including support for Malay and Indonesian. Arrowhead plans to demonstrate the model at Singapore FinTech Festival 2026, scheduled for November 18–20 at Singapore EXPO.

Owning the text-to-speech layer gives Arrowhead control over the details that generic voice systems often treat as afterthoughts: regional speech data, local names, language changes, emotional delivery and response timing. The result is still an automated voice—but one built around how customers in India and Southeast Asia actually speak.

Stay connected via Google News

Follow us for the latest AI News