Arrowhead AI launches Fletch, a voice model built for how India and Southeast Asia actually speak
BENGALURU — Arrowhead AI, a Bengaluru-based voice AI company serving banks, lenders, insurers and consumer businesses, has launched Arrowhead Fletch, an in-house text-to-speech (TTS) model designed for multilingual customer conversations across India and Southeast Asia.
Fletch supports 14 languages, including English, Hindi, Tamil, Telugu, Bengali, Malay and Indonesian. It can switch between languages mid-sentence and adapt its vocal tone to the meaning of a conversation, with styles ranging from empathetic and reassuring to direct and apologetic. The model operates at a reported latency of 80 milliseconds and begins generating speech in under 100 milliseconds.
Unlike generic speech models that may struggle with regional accents, code-mixed sentences and Indian proper nouns, Fletch has been trained on thousands of hours of curated, use-case-specific regional speech, alongside recordings from professional voice artists.
Deployed across live customer calls
Fletch is deployed across live customer calls, supporting conversations across collections, sales, renewals and inbound customer service. Arrowhead AI works with clients across industries including Axis Bank, RBL Bank, IndusInd Bank, Tata 1mg, upGrad, KreditBee, Kissht and TVS Credit, with real-world deployments helping refine the model’s speech generation around terminology, accents and conversational patterns.
According to Arrowhead, the model powers more than one million calls per day across collections, sales, renewals and inbound customer service. These deployments provide real-world use cases for refining speech generation around the terminology, accents and conversational patterns encountered in customer interactions.
The company reports approximately 30% better pronunciation accuracy on selected terms, including brand names, product names, customer names and Indian proper nouns. Arrowhead says the improvement helps make conversations more natural and can reduce instances where customers ask voice agents to repeat themselves.
The pronunciation improvement relates specifically to these terms and should not be interpreted as a 30% improvement in overall call resolution or customer satisfaction.
Building for multilingual customer conversations
In India, customer calls frequently combine regional languages, English and numerical information within the same sentence. A payment reminder, for instance, might combine Hindi and English while communicating an EMI amount and payment date.
Fletch is designed to handle these code-mixed conversations while adapting pronunciation and tone to the context. It supports 12 emotion styles, including calm, apologetic, empathetic, firm, concerned, curious and warm.
By developing its own speech-generation model, Arrowhead can tune pronunciation, latency and vocal expression for its target markets instead of relying exclusively on third-party speech APIs. Fletch operates within the company’s calling platform alongside its speech recognition system, language model and a dedicated model for addresses.
“Owning our voice layer means we can tune every millisecond and every emotion for how India and Southeast Asia actually speak,” said Devyani Gupta, Founder and CEO of Arrowhead AI. “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.”
Vengadanathan Srinivasan, Co-founder and CTO of Arrowhead AI, said, “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.”
Expanding into Southeast Asia
Arrowhead is extending Fletch’s language capabilities to markets beyond India, with Malay and Indonesian support ahead of Singapore FinTech Festival 2026, scheduled for November 18–20 at Singapore EXPO. The company plans to demonstrate the model at booth 6F29.
Technical background
Fletch is built on a DiTAR-based architecture adapted and trained by Arrowhead. DiTAR (Diffusion Transformer Autoregressive Modeling for Speech Generation) combines language modelling with a diffusion transformer to generate continuous speech.
Arrowhead Fletch: Key specifications
●Languages: 14, including English, 12 Indian languages, Malay and Indonesian.
●Time to first audio: Under 100 milliseconds.
●Reported latency: 80 milliseconds.
●Emotion styles: 12.
●Training data: Thousands of hours of curated, use-case-specific regional speech.
●Live deployment: More than one million calls per day, as reported by Arrowhead AI.
●Pronunciation accuracy: Approximately 30% improvement on selected names and terms, as reported by Arrowhead AI.
****
