The Voice AI Experience in India Sucks. But Vobiz.ai Has a Plan to Fix it

The real challenge in building layered AI systems is not the technology itself, but the infrastructure that sits beneath it. One of the critical bottlenecks in model deployment is telecom latency—the difference between sending a data request over a network and receiving a response.

In AI systems, it lives in that gap between what the user feels and what the GPU does. But voice AI systems are far more complex. 

A typical AI voice interaction involves telecom routing, speech-to-text conversion, language model processing, and text-to-speech output. Each step adds a delay.

“Telecom latency is around 500 milliseconds… STT (speech-to-text ), LLM (large language mode), and TTS (text-to-speech) add roughly 800 milliseconds,” explains Vikash Srivastava, Co-founder and CTO of Vobiz.ai. “This becomes abnormal… 1.2 to 1.5 seconds of latency, where it starts feeling like AI is responding slowly.”

That bottleneck is becoming increasingly visible as voice AI adoption moves from experimentation to production. Across sectors such as BFSI, logistics, and healthcare, companies are beginning to rely on AI agents for high-volume customer interactions. Last year, Bajaj Finserv had disbursed over ₹1,000 crore in consumer loans across India through its voice AI agents.

“If you really see, when comparing to 2025 and 2026, telecom latency grew between 300ms and 400ms,” Vobiz.ai Co-founder and CEO Suman Gandham notes. 

Vobiz.ai is trying to solve just that. The Bengaluru-based startup, founded in 2025 and operating with a team of under 50, is focused on the telecom infrastructure that carries every AI conversation.

Earlier, in May, it raised $1 million in a seed round led by Piper Serica VC Fund.

Inside Vobiz, that shift is reflected in its own growth. “We were doing around 30,000 calls per day. Right now, we are doing around 3 million calls per day… in less than one year,” he said.

The Problem Statement

In a typical voice AI call, three core layers sit between what a user says and what the AI replies.

STT converts the caller’s voice into text so it can be understood by the system. That text is then processed by an LLM, which interprets intent and generates a response. Finally, TTS converts that response back into audio for the user. Each of these steps happens in sequence and adds latency, which is why even small delays across STT, LLM, and TTS can make a conversation feel slow or unnatural in real-time.

While AI models are evolving rapidly, telecom infrastructure has remained largely unchanged. As Srivastava now puts it, “In the STT, LLM, and TTS world, there is constant evolution… Telecom still sits at the same 500 milliseconds. There is no evolution going on there.”

This gap becomes most apparent outside controlled environments. Early pilots often perform well in office settings, but break down in real-world conditions. “In POCs… you are sitting in an office, talking clearly,” Gandham explains. “But in real life… in the middle of the road… all this noise, latency… happens.” 

Noise, echo, and unclear speech can distort inputs, while long conversations can cause AI systems to lose context. “If your voice is not clear… AI will assume something else and respond,” Srivastava adds.

India’s linguistic diversity further complicates the problem. Accents and languages shift rapidly across regions, while datasets for many of these languages remain limited. This slows the development of robust voice systems, even as demand accelerates.

AI-Native Telephony

It was this gap, between AI capability and real-world performance, that led the founders to reframe the problem. Instead of improving the AI layer, Vobiz.ai set out to rebuild the telecom layer underneath it. 

“We are a media processing unit… to give a very clean outcome to the AI layer,” Srivastava shares. 

The system handles real-time audio streaming, noise reduction, transcription, answering machine detection, and the redaction of sensitive information. At its core, the architecture is designed to reduce latency by minimising the number of intermediary steps in a call. “We ensure we have a single hop operation… Every hop… adds anywhere between 20–60 milliseconds,” he explains.

Beyond performance, regulation is another constraint shaping the space. Telecom in India is tightly governed, and compliance failures can lead to service disruptions. 

Vobiz is attempting to simplify this layer by automating onboarding and verification processes through integrations with PAN, GST, and Aadhaar systems. However, Gandham acknowledges that “regulation still works in a very stringent and old school way.”

Scaling Voice Infrastructure

The company’s customers are primarily businesses building voice AI systems, including Sarvam, Razorpay, Shiprocket, Haptik, and others. 

Gandham notes that Vobiz is growing 80% month-on-month and increasingly focusing on enterprise contracts. He positions the company against its competitors, legacy CPaaS providers, which he describes as slower and more manual, with onboarding processes that can take weeks.

“Traditional CPaaS stacks were designed to route calls from point A to point B, not to serve as real-time infrastructure for software making decisions mid-conversation… developers glue together three or four services just to get a call working, and end up with 1,400–1,700 ms of latency that makes the agent sound broken,” Srivastava earlier explained in the funding announcement.

The funding reflects investor conviction in this thesis. “In Vobiz, we're excited to back a team that spotted the infrastructure problem with voice AI before the market did,” said Abhay Agarwal, Founder of Piper Serica. “They know exactly where it breaks, and why patching it doesn't work.”

Vobiz.ai will use the capital to strengthen engineering and go-to-market efforts as the company works towards a $5 million ARR target. It also plans to expand its carrier partnerships, improve real-time streaming capabilities, and introduce advanced observability and compliance tools over the next 12–18 months.

However, the race to build voice infrastructure is no longer limited to India. Earlier in May, Krisp, a Berkeley, California-based voice AI company known for its real-time noise cancellation and audio processing technologies, announced VIVA 2.0, a new infrastructure layer for voice AI agents focused on interruption handling, accent detection, conversational turn-taking, and real-time audio enhancement before audio even reaches the STT and LLM layers. 

Alongside infrastructure players such as Twilio and Telnyx, and newer voice AI platforms like Vapi, Retell AI, and LiveKit, a broader category is beginning to emerge globally around AI-native voice infrastructure rather than just voice agents themselves. 

The implication is clear: voice AI is no longer just about STT, LLMs, and TTS models. It is about everything that happens before and between them—latency, audio quality, interruptions, routing, compliance, and telecom reliability in real-world environments.