Shunya Labs wants voice infrastructure to work the way Indians speak

Most voice AI systems are trained for a speaker who rarely exists in India: someone speaking unaccented English.

For Ritu Mehrotra and Sourav Bandyopadhyay, that gap became impossible to ignore while building United We Care, a company focused on mental health. 

Users switched between languages mid-sentence, spoke in dozens of accents, used specialised medical terminology, and raised concerns about sensitive data being routed through third-party servers, while the existing voice infrastructure struggled to keep up.

“When we looked at the market, we realised nobody had built models that could truly handle the diversity of speech patterns across Asia,” Mehrotra says.

Rather than working around those limitations, the founders decided to build the infrastructure themselves. That became Shunya Labs.

Today, the company builds proprietary voice AI infrastructure designed for multilingual, accented, and code-switched speech. Its technology powers enterprise customers across banking, healthcare, telecom, automotive, and ecommerce, while running on CPUs rather than expensive GPU-heavy deployments.

Mehrotra, an IIFT and Harvard Business School alumna, previously led growth and operations at Booking.com and Zomato. Bandyopadhyay is an electronics engineer from NIT Jamshedpur and a PhD researcher at IIT Kharagpur. 

They were joined by Abhishek Sharma, who previously served as COO of Dineout and CHRO at Times Internet.

Building the stack

Shunya Labs has built its voice AI stack from the ground up.

The platform comprises three core layers: a speech-to-text model that converts spoken language into text, an orchestration layer that processes the information with context and memory, and a text-to-speech model that converts responses back into voice.

The company says the entire stack can run on a CPU, whether on a phone, inside a vehicle, or on a company's own servers, with or without internet connectivity. 

According to Shunya, this reduces deployment costs to roughly one-twentieth of comparable providers while covering more than 220 languages, including 55 Indian languages.

The models are proprietary, but enterprise customers get an additional custom layer trained on business-specific vocabulary, product names, and terminology, telecom plan names, pharmaceutical terms, banking product catalogues, and the like. Building one of these custom models takes about 40 GPU hours, and adding a new language takes around 100, against an industry norm closer to 10,000.

Before audio reaches the transcription engine, a proprietary denoiser removes background noise while another tool corrects volume, frequency, and distortion. The system is designed for environments where audio quality is inconsistent, including call centres and field recordings.

On top of transcription, the platform offers sentiment analysis, emotion tracking, intent detection, and speaker identification.

Mehrotra says Shunya's Hinglish model can return mixed-language tokens in real time rather than forcing users into a single language.

In February 2026, the company launched Vāķ, an open-weight real-time translation system spanning 55 Indian languages, at the India AI Impact Summit in New Delhi.

Built with Nasscom, whose GenAI cohort Shunya is part of, Vāķ offers less than 1.5 seconds of end-to-end latency, zero-shot voice cloning, and runs on the company's Zero Suite. Model weights are available on Hugging Face.

Bandyopadhyay describes Vāķ as “sovereign, self-funded innovation”, adding that the goal is for “every developer can build, every government can deploy, and every Indian can be heard in their own language”.

Winning enterprise customers

Shunya works exclusively with large enterprise customers.

The sales process follows three stages: a proof of concept under NDA, a pilot in a single region and language, and finally a full deployment that typically takes six to eight months before converting into a recurring contract.

The company recently shifted its proof-of-concept stage from free to paid.

Pricing follows a three-tier model. Individual developers can access the platform for free or purchase prepaid usage credits starting at $500. Enterprise customers receive custom pricing, self-hosted deployment options, dedicated service-level agreements, and higher concurrency support.

According to the company, a typical voice agent combining speech-to-text, an LLM, and text-to-speech costs about $0.0154 per minute to run.

Mehrotra says revenue has doubled month on month over the past three months, while platform usage has grown 6-8X over the same period.

The company also points to top-ranked automatic speech recognition performance across multiple public benchmarks on Hugging Face, including finance, TED Talks, and European language datasets, where it says its models outperformed offerings from Nvidia and Meta.

Among the metrics cited by the company, Zero STT Med records an 11.1% word error rate and a 5.1% character error rate on noisy, multi-speaker medical audio. Zero Tinny ONNX reports a 3.10% word error rate, while Zero STT Indic claims up to 200% better accuracy than the next-best model for Indic speech.

Where the technology is used

The platform is ISO 27001 certified, HIPAA and GDPR compliant, and can be deployed across cloud, edge, and on-premise environments.

Shunya's customer base spans telecom, banking and financial services, healthcare, automotive, and ecommerce. Customers include Panasonic, NASSCOM, Gomotto in the US, OMG Pharma in Australia, and ZET in India.

In banking, the platform handles KYC verification, payment reminders, and mid-call language switching. Healthcare deployments transcribe clinical conversations in real time, separate doctor and patient voices, and structure outputs for electronic medical records while remaining on-premises.

Contact centres use the platform for live transcription and real-time agent assistance. In automotive applications, the models run within infotainment systems without requiring a network connection. Ecommerce customers use the technology to convert spoken requirements into relevant product results, which Shunya describes as voice-to-intent.

The company's work is supported by 12 patents covering areas including clinical translation, hallucination mitigation, and text-to-emotion mapping. It has also produced more than 18 peer-reviewed publications and holds 23 world records. Several models are available as open source on Hugging Face.

Looking ahead

India's conversational AI market is projected to reach $3.7 billion by 2033, according to Grand View Research, growing at a CAGR of 26.4% from 2026.

The market already includes players such as Sarvam, Gnani.ai, and Yellow.ai, which operate across different layers of the voice AI ecosystem.

Shunya has raised about $5.5 million across pre-seed and seed rounds and plans to raise a Series A round within the next three to four months. For now, however, profitability is the priority.

“In the next six to eight months, the primary goal is to reach P&L positivity, with the focus on building revenue infrastructure while keeping costs efficient,” Mehrotra says.

The company says it is not pursuing a burn-and-scale strategy and is instead focused on building sustainable recurring revenue through enterprise deployments across telecom, BFSI, healthcare, and automotive markets in Asia.

Its longer-term ambition extends beyond India. “Our target population includes people across Asia, Eastern Europe, and other regions with high linguistic diversity and lower digital literacy,” Mehrotra says.

She argues that someone who cannot type in English, or cannot type at all, can still speak in their own language. The company's bet is that these users need foundational voice infrastructure built for them, rather than systems designed around English-first interfaces and abundant computing power.