Most voice artificial intelligence systems are trained on samples of native speakers of flawless English, which rarely reflects the reality in India. However, for Ritu Mehrotra and Saurav Bandopadhyay, founders of United We Care, a mental health organization, this gap became apparent.
Users often switched between languages mid-sentence, spoke with numerous accents, used specialized medical terminology, and expressed concerns about transmitting confidential data through third-party servers. The existing voice infrastructure failed to meet these requirements.
Mehrotra noted that market research revealed no one had created models capable of adequately processing the diversity of speech patterns across Asia. Instead of circumventing these limitations, the founders decided to develop the necessary infrastructure themselves, leading to the creation of Shunya Labs.
Today, the company is building its own voice AI infrastructure designed to handle multilingual, accented, and code-switched speech. This technology is utilized by large corporate clients in the banking, healthcare, telecommunications, automotive, and e-commerce sectors, all running on central processing units (CPUs) rather than expensive solutions requiring powerful graphics processing units (GPUs).
Shunya Labs built its voice AI stack from scratch. The platform consists of three core components: a speech-to-text model, an orchestration layer for context and memory-aware information processing, and a text-to-speech model for vocalizing responses.
The company asserts that this entire stack can operate on a CPU, whether used on a phone, in a car, or on the company's own servers, and even without an internet connection. This allows deployment costs to be reduced by approximately twenty percent compared to similar providers, while supporting over 220 languages, including 55 Indian languages.
The models are proprietary, but corporate clients receive an additional customizable layer that trains on business-specific vocabulary, product names, telecom plan terms, pharmaceutical names, and bank product catalogs. Creating such a custom model takes about 40 hours on a GPU, and adding a new language requires approximately 100 hours, significantly less than the industry standard of 10,000 hours.
Before the audio signal enters the transcription engine, a patented noise suppressor eliminates background noise, and another tool adjusts volume, frequency, and distortion. The system is designed to function under conditions of unstable sound quality, such as in call centers or field recordings.
In addition to transcription, the platform provides features like sentiment analysis, emotion tracking, intent detection, and speaker identification. Mehrotra emphasized that Shunya's Hinglish model can output mixed-language tokens in real time, without forcing users to limit themselves to a single language.
In February 2026, the company unveiled Vāķ—an open-weight, real-time translation system covering 55 Indian languages—at the India AI Impact Summit in New Delhi. Vāķ was developed in collaboration with Nasscom, which includes the Shunya GenAI cohort. The system achieves end-to-end latency of less than 1.5 seconds, supports zero-shot voice cloning without prior training, and operates on its proprietary Zero Suite platform. Model weights are available on Hugging Face.
Bandopadhyay describes Vāķ as a 'sovereign, self-sustaining innovation,' adding that the goal is for 'every developer to be able to create, every government to be able to implement, and every Indian to be heard in their native language.'
Shunya exclusively collaborates with large corporate clients. The sales process involves three stages: proof of concept (PoC) under a non-disclosure agreement (NDA), pilot testing in one region and one language, and finally, full-scale implementation, which typically takes six to eight months before transitioning to a permanent contract.
The company recently changed its policy to make the proof of concept stage paid. The pricing structure is implemented in a three-tiered model. Individual developers can use the platform for free or purchase prepaid usage credits starting from $500 USD. Corporate clients receive custom pricing, self-hosting options, dedicated service level agreements (SLAs), and higher concurrency support.
According to the company, a typical voice agent combining speech-to-text, LLM, and text-to-speech costs about $0.0154 per minute of operation. Mehrotra reported that the company's revenue has doubled monthly over the last three months, and platform usage has increased 6–8 times during the same period.
The company also points to high automatic speech recognition scores in various public benchmarks on Hugging Face, including financial, TED Talks, and European language datasets, where its models outperformed offerings from Nvidia and Meta. Among the metrics, Zero STT Med demonstrates a word error rate of 11.1% and a character error rate of 5.1% when handling noisy medical audio with multiple speakers. Zero Tinny ONNX shows a word error rate of 3.10%, and Zero STT Indic claims an accuracy 200% higher than the next best model for Indian speech.
The platform is certified according to ISO 27001, complies with HIPAA and GDPR requirements, and can be deployed in cloud, edge, or on-premise environments. Shunya's client base spans telecommunications, banking, financial services, healthcare, automotive, and e-commerce. Clients include Panasonic, NASSCOM, Gomotto in the US, OMG Pharma in Australia, and ZET in India.
In banking, the platform is used for KYC verification, payment reminders, and language switching during calls. In healthcare, it transcribes clinical conversations in real time, separates the voices of the doctor and patient, structuring results for electronic health records while keeping data local. Call centers use the platform for live transcription and real-time agent assistance. In the automotive industry, the models operate in infotainment systems without needing network connectivity. E-commerce clients use the technology to convert spoken queries into relevant search results, which Shunya calls voice-to-intent.
The company's work is supported by 12 patents covering areas such as clinical translation, hallucination mitigation, and text-to-emotion mapping. Furthermore, over 18 peer-reviewed papers have been published, and the company holds 23 world records. Many models are available as open source on Hugging Face.
According to Grand View Research, the conversational AI market in India is projected to reach $3.7 billion by 2033, demonstrating a Compound Annual Growth Rate (CAGR) of 26.4% starting from 2026. Players such as Sarvam, Gnani.ai, and Yellow.ai are already present in the market, operating at different levels of the voice AI ecosystem.
Shunya has raised approximately $5.5 million in pre-seed and seed rounds and plans to conduct a Series A round within the next three to four months. However, the current priority is achieving profitability. Mehrotra stated that the main objective over the next six to eight months is to achieve a positive P&L, focusing on building a revenue infrastructure while maintaining cost efficiency.
The company does not follow a 'burn and scale' strategy; instead, it focuses on building sustainable recurring revenue through corporate implementations in the Asian telecommunications, BFSI, healthcare, and automotive sectors. The company's long-term ambitions extend beyond India. Mehrotra indicated that the target audience includes people across Asia, Eastern Europe, and other regions with high linguistic diversity and low digital literacy. She believes that people who cannot type in English or cannot type at all can still speak their native language, and these users need basic voice infrastructure specifically designed for them, rather than systems oriented towards English-dominant interfaces and heavy computational power.
Despite the development of artificial intelligence technologies, employees continue to prefer voice communication methods, especially when quick response to emerging issues is required. According to a recent Mitel report titled 'State of Workforce Communication,' 79% of employees stated that they prioritize voice communication for prompt coordination of actions, using messengers or video calls only for subsequent clarification.
When choosing a communication method for urgent or critical moments, voice remains the most popular choice across all age groups. This preference ranges from 43% among Generation Z workers to 54% among Baby Boomers. This underscores the enduring significance of real-time communication in high-risk work environments where delays can negatively affect service, safety, or operational performance.
The study, conducted by Vanson Bourne on behalf of Mitel, surveyed IT managers, office workers, and frontline staff to better understand the effectiveness of communication tools across different roles and working conditions. Special attention is paid to the fact that eight out of ten (80%) workers globally operate on the front lines, including healthcare, manufacturing, retail, hospitality, financial services, and the public sector, where communication gaps can have immediate and significant consequences.
As workplace technologies modernize and AI-based tools are implemented, progress assessment must consider how these investments simplify communication in real-world settings, whether they align with the workforce profile, and if they contribute to achieving desired outcomes directly at the point of task execution.
Many frontline employees, despite their central role in providing services and conducting operations, operate in communication environments that were not originally designed for mobile, time-sensitive, and field work. This often means having to search for information in multiple locations and an increased risk of messages being missed at crucial moments.
Office and frontline workers use an average of seven communication channels, leading to experience fragmentation. This makes it difficult to find the necessary information, contact the right person, or get a quick response. Among frontline staff, 65% report time loss due to switching between different tools, and 64% report missed messages due to their distribution across multiple channels.
These issues affect not only productivity. Mitel found that 70% of frontline workers believe that complex or unreliable communication tools hinder them from providing the best support to customers, patients, or clients. More than a third of respondents claim that communication problems create risks to the safety of customers, patients, or personnel.
Martin Bitzinger, Senior Vice President of Product Management at Mitel, stated: 'The future of workforce communications will not be exclusively voice, but it must be voice first in moments when real-time coordination is most critical.' He added that for workers in critical industries and frontline staff, voice is often the fastest, clearest, and most reliable way to coordinate when work cannot wait. He also noted that when modernizing and implementing AI, organizations must not forget to create a flexible communication infrastructure that supports employees regardless of their role or function in task execution and outcome achievement.
Healthcare serves as a prime example. In dynamic clinical settings, workers need to quickly coordinate actions between care teams, different locations, and devices. Communication failures can lead to delays in assistance during critical moments, increasing patient waiting times for appointments or discharge and reducing the hospital's capacity to serve subsequent patients.
For healthcare leaders, these risks are not abstract; they are directly linked to patient satisfaction, operational costs, and safety outcomes. The Mitel study showed that healthcare is one of the industries where communication problems are most likely to pose safety threats, and frontline workers in this sector report the most acute gaps in having their needs considered when making decisions about communication technology and infrastructure. These difficulties extend to other critical industries, including manufacturing, financial services, hospitality, and retail.
Zeus Kerravala, founder and lead analyst at ZK Research, emphasized: 'In a crisis situation, the interface must disappear. Voice does that.' He explained that you cannot train a nurse or dispatcher to navigate a user interface under pressure—you need a channel that works when needed. This is why voice remains the only modality that stands up to the test when stakes are high.
As AI adoption grows across all business sectors, voice is also becoming the most natural and effective way for workers to interact with technology. Especially for frontline staff, a voice-oriented AI interface can increase user adoption and productivity by reducing the time spent navigating menus, switching between applications, or typing text messages.
Workforce technology development strategies should not view voice as an outdated or backup channel; it must become an integral part of a modern communication strategy. This allows voice to act as an interface layer through which AI provides the greatest value to the workers who need it most.
Luis Domingos, CTO of Mitel, stated: 'AI has the potential to radically impact those workers who have been least provided for by digital transformation, but we must meet them where they are.' He concluded that whether it is a nurse in the emergency room or a technician in the field, voice AI bridges the gap in a way no previous technology has. In his view, the application of the future, especially for frontline workers, is not a chat window or a dashboard, but a dialogue.
