Nikhil PahwaonSeptember 3, 20265 minute read

More human than human?

Note: This article was originally published on Reasoned and is being cross-posted on MediaNama. Read the original version here.

At a recent presentation on Voice AI, the speaker showcased an interaction with an AI agent that the system showed as a successful call: the Voice AI agent heard the caller and continued with its script of querying the customer, while the recipient on the other side kept saying “Hello?” to everything the agent said.

On the face of it, Voice AI relies on two basic technologies:

Speech To Text (STT): Neeraj Arora, now CEO for India, MENA and Managing Director at General Catalyst, once told me, while we were on a flight to Pune for Proto. in, that speech-to-text is the hard problem for the world to solve. He was with Google then, looking at M&A in India for the company. That was over a decade and a half ago, and it has largely been solved now, especially for Indian users and dialects. A historical anecdote: in 2008, Google launched a toll-free number 180041999999, where users could call in and talk to human operators for search queries. In 2010, Google launched a Mobile app, “which enables users to search by voice, and is compatible with a variety of Indian accents”, because the voice search using the toll-free number “helped us learn more about user needs and build core technologies like speech recognition, which are helping change how people search and perform other activities on the web.” The toll-free number was shut down in 2010, after they had collected enough information. Speech-to-text forms the foundation of most Voice AI services, whether it is note-taking services like Granola and Letterly, or voice input services like Wisprflow. OpenAI’s Whisper is open-sourced.

Text-to-speech (TTS) has been a solved problem for a long time, but if you see the tooling that exists for speech-to-text, the problems that remain are twofold: one is that many voices sound robotic, though there are still many that sound human. You can use free a text to speech models like Piper, which sounds a bit robotic, or Kokoro, which is better.

The challenges that remain with Voice AI

Things when Voice AI has to become interactive. Most voice deployments are taking place in the call center business: for a human being to handle a call, costs can be as high as Rs. 40 per call, and the longer the call goes on, the more it costs the call center per minute. So the cost of Rs. 2 per call, that has recently led to some controversy, is genuinely useful for businesses: the math works out, and businesses factor in some amount of failure here.

The mechanism for most voice deployments follows a variation of the following sequence:

Listen to the customer —> Transcribe the text (STT) —> Run it through a reasoning model, which drafts a response —> Speak (TTS) —> Listen to the customer

In deployment, the following issues emerge (based on my experience and that of the people who spoke at the discussion):

One, the latency between the consumer speaking and the agent responding can be disconcerting: if your deployment isn’t backed by enough compute, the processing cycle above can take far too long, which leads to customer dissatisfaction.

Two, building the system for interruptions is complicated: A customer can interrupt a conversation midway through an AI response, and that starts a whole new processing cycle. If your response is AI-generated verbal diarrhea, the customer will interrupt. Humans on a call know how to deal with interruptions, can sense a customer’s irritation/lack of patience, and can navigate this scenario better.

Three, pauses do not necessarily indicate completion: A customer can pause during a conversation, sometimes trying to recall information, and a human would know whether to wait or not. An AI tool can treat this as the completion of an input and start responding.

Four, background noise is not an input: a demo environment is sanitised, but in the real world, a shopkeeper on a customer care call has people speaking all around him. She might pause to give someone else an instruction, but an agent can treat ambient conversations as input, or a comment meant for someone else as an input for the agent. Interruption handling is largely unresolved with Voice AI.

Five, connections are patchy: connectivity over mobile can be patchy for the agent, which means there are gaps in sentences, or agents need to figure out what exactly someone meant to say when the connection is patchy.

Six, AI agents can still mess up grammar and names: A female voice being chosen to engage with customers is still voicing an AI-generated script, and could be speaking with male grammar in the response. In English, the grammar is largely gender neutral, but gender is an integral part of Indic languages. Most demos are in English or Hindi, but as you move to regional languages, grammar mistakes blow up. The same name can be pronounced differently in different regions, and the AI model can mix those up.

Seven, in outbound, the standard opening can lose the customer: The standard approach to voice scripts is to first introduce yourself, which is something that call center agents do. A person who deployed Voice AI said that what worked for them instead was to cut to the chase, seek requirements or directly get inputs from the customer, instead of an AI-generated introduction that the customer has no patience for. The first 10-15 seconds, as with any YouTube video, are critical.

Eight, the math doesn’t work out as pitched: While the talk is of Rs. 2 per minute, an agentic 10-minute call can cost close to Rs 15 per minute. Customer-support calls generally run longer, so the cost has to be evaluated on the basis of those costs. A speaker said that companies pitch 10x or 20x in terms of cost saving, but the actual value saved is only 3x to 5x, which is substantial, but not as pitched. There is also a cost to setting up and optimising the system, with a team involved in running it. This changes the build-versus-buy decision: a self-serve platform may be useful for testing, but if a company has a technical team, then build gives more control.

Nine, models can change: Reasoning models deprecate over time. They can behave differently in different conditions. Service providers can “upgrade” models, which require changes in how your system works. The systems need to be watched perpetually for changes, and compared with benchmarks, and instead of a call center team, you’d be operating an audit team.

Ten, analytics can fail: A person who spoke at this discussion said that when they ran an outbound dialer campaign, they found that the agent was qualifying the leads incorrectly: people who were interested were marked as not interested.

My takeaway from the discussion was that voice quality is now a relatively small part of the work, and companies need to work on the job to be done and optimise for real-world interaction, including noise and interruptions.

What stuck with me was the idea that we are probably going to reach a point when AI interactions are normalised in terms of customer expectation setting, and so that customers themselves learn how to navigate an AI customer support executive, just like IVR is learned behaviour. We don’t mind robotic-sounding voices when we have learned to expect them.