Bengaluru's Arrowhead AI launches Fletch voice model for 14 languages

Arrowhead AI, a voice artificial intelligence company based in Bengaluru, has launched Fletch, an in-house text-to-speech model built for customer phone calls in India and Southeast Asia, according to reports published between 30 October and 2 November 2026. Text-to-speech, or TTS, is the technology that turns written text into spoken words. The company says Fletch speaks 14 languages, including English, Hindi, Tamil, Telugu, Bengali, Malay and Indonesian, and can switch between them in the middle of a sentence, the way many people in the region actually talk. With Fletch, Arrowhead says it now runs its own speech recognition, speech generation and language model, rather than relying on outside providers for any of them.

Many companies in India now use AI voice agents to make and answer phone calls: reminding customers of loan payments, confirming deliveries, handling insurance queries or selling products. These systems need three parts. Speech recognition turns the caller's words into text, a language model decides what to say, and text-to-speech speaks the reply. If the voice sounds robotic, mispronounces a name or pauses too long, callers hang up or lose trust.

India makes that harder than most markets. People routinely mix languages, such as Hindi and English in the same sentence, a habit known as code-mixing. Names of people, places and products come from many languages. Global voice models trained mostly on English can stumble on both. Arrowhead's pitch is that a model trained on regional speech can sound natural to Indian callers. Its claimed figures, such as the 80-millisecond latency and better pronunciation, come from the company and have not been independently tested.

Owning the whole "voice stack" also matters for business. Paying an outside provider for each minute of speech adds cost at scale, and sends customer data to a third party. A company that runs its own models can control cost, latency and data handling, which is important for banks and insurers with strict privacy rules.

Expansion into Malay and Indonesian points to Southeast Asia, where similar multilingual and code-mixed speech is common. Arrowhead is adding that support ahead of the Singapore FinTech Festival 2026, according to reports.

Peepals calculation: If the system handles one million calls a day, that is about 694 calls a minute on average (10,00,000 ÷ 1,440 ≈ 694), or more than 11 a second (694 ÷ 60 ≈ 11.6). A latency of 80 milliseconds is less than a tenth of a second (80 ÷ 1,000 = 0.08). Our calculation from the reported figures.

Short-term: Arrowhead's business customers can use Fletch for multilingual calls, and the company can offer a complete in-house voice system.

Long-term: If home-grown voice models prove better at Indian speech than global ones, more Indian companies may build or buy local AI rather than rely on foreign providers.

Who is affected: Banks, insurers, lenders and e-commerce firms that use AI calling; customers who receive such calls; and Indian AI developers working on regional-language speech.

Questions and Answers