I had to change my ChatGPT voice to the annoyed British guy. I’m scared that if the voice is too friendly I’ll fall in love with it.
Like that guy. In that movie.
Let’s talk about voice assistants.
Siri used to be the butt of the joke. But while we were busy asking Siri how to hide a body, voice AI quietly permeated all corners of the market. As of 2025, 67% of organizations consider voice AI as being core to their business.
Those organizations realize that AI agents are better with speech capabilities.
Oh, and that movie I referenced? Not such a far cry. Open AI’s recent acquisition of io is expected to be with the intention of building a non-invasive, perpetually aware voice assistant.
You know, a little buddy in your ear at all times.
So here we are: Alexa is more recognizable as a product than as a person’s name, AI companies’ CEOs are taking engagement photos together, and two thirds of businesses have already saved-the-date.
And if you’re not on top of it, then sister, you’re behind.
Which is understandable. The technology is enigmatic, and there aren’t a whole lot of folks explaining how it works. But guess who has two thumbs and a graduate degree in speech technology?
(You can’t see but I’m sticking up my thumbs.)
(...You know who else can’t see? Voice assistants.)
(I digress.)
I’m writing this article to catch you up to speed. We’ll talk about AI Voice Assistants: how they work, what you can do with them, and the reasons so many companies are opting to integrate them into their operations.
What is an AI Voice Assistant?
An AI voice assistant is AI-powered software that processes speech input, understands it, executes tasks, and provides responses to the user. These assistants are used across industries and use cases, adding a personal touch to task management and customer support.
How do AI Voice Assistants Work?
AI voice assistants are a complex orchestration of AI technologies. In the few seconds between capturing the user’s input speech and generating a response, a number of processes are triggered to deliver a seamless interaction.
Automatic Speech Recognition (ASR)
Automatic speech recognition is sometimes called speech-to-text, because that’s what it is.
When a user speaks into their device– be it a phone, home assistant, or car dashboard, their speech is converted into text. To do this, deep neural networks are trained to predict the transcription of an audio clip.
After training on 1,000s of hours of speech data across millions of different clips involving different speakers, accents, and noise conditions, these AI models get pretty good at transcribing.
And that’s important– the first step in the multi-layer system needs to be robust.
Natural Language Processing (NLP)
With the speech input transcribed, the model moves on to interpreting it.
NLP is the umbrella concept for all the techniques used to parse the user’s query (as transcribed text) into intent and meaningful units.
Intent Recognition
Text is unstructured, and the task of teasing out meaning is far from trivial. Take the following few queries:
- “Schedule a call with Aniqa for Tuesday at 1.”
- “Can you play Cher?”
- “What goes well with goat cheese?”
An AI assistant will have a finite series of intents under the hood. For our bot, that would include:
- booking appointments
- playing media
- possibly searching the web, and
- casually conversing
Intent recognition is responsible for classifying each user query into one of these categories.
So, which one does each of our examples fall under?
“Schedule a call…” is phrased as an imperative. Relatively straightforward. “Can you…?” is phrased as a question. But it’s also a command, like the previous query. In both cases, you intuitively understand the desired action, but it’s not so easy to formalize.
“What goes well with…?” is simple– sort of.
We know what kind of answer we want: food. But it’s not super clear where it should grab the answer from.
Should it search the web? If so, how many responses should it give? The first result wouldn’t be very thorough, but giving a lot of responses can overcomplicate a simple task.
On the other hand, maybe it can just dig from its internal knowledge– but we’re getting ahead of ourselves.
The takeaway is: the choice isn’t always simple, and the complexity of this task has as much to do with the design– or personality– of the bot as it does with the user’s query.
Named Entity Recognition
Above and beyond knowing which task to perform, the bot needs to recognize the provided information.
Named entity recognition is concerned with extracting the meaningful units – or named entities – from unstructured text. For example, identifying the names of people’s names, musical artists, or dates in a user’s query.
Let’s have a look at the first query again:
- “Schedule a call with Aniqa for Tuesday at 1.”
Aniqa is a person, and it’s implied from the query that the user knows her. That makes her– in all likelihood– a contact.
In this case, “contact” would be pre-programmed as an entity, and the bot would have access to the user’s contacts.
This goes for times, locations, and any other meaningful information that might be hiding in a user query.
Retrieving Information
Having understood what you want, the voice assistant has to search for relevant information to help it respond. A good bot will be equipped with a whole suite of extensions to help meet your needs.
We talked about internal knowledge earlier. I’m sure you were blown away at some point by large language models’ (LLM) and their extensive knowledge. And it’s impressive, but as your queries get more specialized the cracks start to show.
Retrieval-Augmented Generation (RAG)
A good assistant has access to external knowledge sources – it doesn’t rely solely on the knowledge it’s acquired during training. RAG conditions the AI’s responses on that knowledge.
Knowledge, in this case, refers to documents, tables, images, or basically anything that can be digitally processed.
It searches through the documentation, pulling items that are most relevant to the user’s query and using them to inform the model’s responses.
- Sometimes it’s in the interest of sharpening an LLMs information, like having it reference academic literature when doing research.
- Other times it’s about giving access to information that the model wouldn’t otherwise have, like customer data.
In either case, it has the added advantage of citing its sources, making responses more reliable and verifiable.
APIs and Integrations
In the same way that an LLM can interface with external information, APIs and integrations allow it to interface with external technologies.
Want to book a Google Meets appointment via Calendly to follow up on a HubSpot lead evaluated with Clearbit enrichment? Unless you built the calendar, video conferencing technology, CRM, and analytics tool (which is highly inadvisable), you’ll need to 🔌integrate⚡️.
These 3rd party tools usually have APIs exposing operations so that they can be performed by other automated technologies– like your agent.
Integrations make it even easier for a bot to interface with 3rd party technology. It’s built on top of an API, covering the messy so you can hook your agent up with little work.
Responding and Text-to-Speech (TTS)
So, the user input’s been transcribed, their intent parsed, the relevant information retrieved, and the task has been executed.
Now it’s time to respond.
Whether it’s answering the user’s question or confirming that it performed the requested task, a voice bot pretty much always offers a response.
Text-to-Speech (TTS)
Equal and opposite to speech recognition is speech synthesis, or text-to-speech.
These are models, also trained on speech-text pairs, often conditioned on speaker, intonation, and emotion to deliver a human-like utterance.
TTS closes the loop that begins and ends with human(-oid) speech.
The Benefits of Voice Assistants
A voice layer on top of AI’s functionality improves the experience all around. Sure, it’s personalized and intuitive, but it’s got advantages on the business side of things, too.
Voice is Faster than Text
With the proliferation of chatbots, users have gotten accustomed to quick responses. With voice AI assistants, we’ve also managed to improve input time.
Voice AI agents keep us from having to formulate proper sentences. Instead, you can blurt out a stream-of-consciousness, and have the bot understand it.
Same goes for the responses. I’ll be the first to admit that reading can be a drag– but it’s not a problem when the responses are narrated to you.
24/7 Responses
Yet another kind of speed. With people working remotely, and business transactions happening across continents, it’s impossible to account for all the timezones and working hours you’ll need to cover.
Spoken interactions should be available to everyone, not just customers who fall into certain working hours. And with voice AI assistants, that could be a reality.
More Personalized Interactions
Talking is about so much more than words. Having a voice bot creates a more personal experience that instills a sense of confidence in the user. Coupled with AI chatbots’ human-like qualities, a voice layer makes for a stronger connection.
Easy Integrate
The fact that voice assistants are hands-free means they’re also UI-free. They don’t require screens, or use of your eyes– which is why they’re so popular in cars.
In fact, they can integrate anywhere that a microphones can be hooked up. That’s a very low bar to cross, not only because microphones are so small, but because they’re everywhere already: computers, smartphones, and even landlines.
Name another cutting-edge technology that’s accessible via rotary telephones.
More Accessible
“Hands-free” isn’t only about convenience. For people with diverse needs, it can be a necessity.
Voice assistants are available to people with mobility-, vision-, and literacy-diversity who might otherwise struggle with traditional AI interfaces.
Use Cases of Voice Bots Across Industries
So, you’re sold on voice bots. Great. But how do you put them into use?
Well, the good news is that pretty much every industry can be improved with voice AI.
Healthcare
Healthcare procedures are notoriously tedious. And for good reason: it’s high stakes work, and it has to be done right. This space is begging for AI automation, provided that it’s reliable and effective.
We’re seeing applications of AI in healthcare already, and voice adds a slew of new opportunities to improve.
A great example of this would be medical questionnaires: personal information, medical history, etc.
Those are tedious. But they’re important.
The gains in speed and productivity alleviate the workload of overworked healthcare professionals, and the human-like conversation flow breaks up the monotony of answering question-after-question.
Accessibility is accounted for, and per the vigorous, multi-layered pipeline we discussed earlier, I can assure you the technology is reliable.
Banking
Speaking of high-stakes and tedious.
Things like checking account balances and updating information are relatively simple transactions, but have a couple of layers of safeguards to reduce errors and fraud.
NatWest’s voice agent deals with regular transactions, freeing up human agents to spend more time on sensitive or complex interactions, driving up customer satisfaction by 150% without compromising on security.
Customer Support
On the topic of automating routine calls, Vodafone’s SuperTOBI, a voice AI assistant, has improved their net promoter score (NPS) from 14 to 64.
That’s because custer service interactions are repetitive, and customers’ queries are answered all the same, whether by a person or an agent. This approach doesn’t compromise on edge cases– those are handed off to human agents.
Retail
I kind of miss the days of speaking with a salesperson.
The problem is, they’re too busy to familiarize themselves with the store’s catalogue and policies, not to mention the time it takes to deal with every individual client.
Enter voice sales assistants like Lowe’s’ MyLow: a virtual sales associate with information on product details, inventory, and policy.
LLMs’ generalized knowledge really shines here: beyond giving Lowe’s-specific information, it uses interior design knowledge to advise customers on home decorating.
Some customers are still looking for human interaction. Fortunately, MyLow is also available to sales associates. Employees can grab the information they need from MyLow and help the customer themselves.
Start Offering AI Voice Assistants
Voice AI assistants are the clear way to go. Efficiency and personality, without compromising on humanity– it’s a win-win.
Botpress offers a customizable drag-and-drop builder, human in the loop oversight, a host of pre-build integrations, and to top it off, a voice wrapper that sits seamlessly atop your agent.
Our bots are clean and intuitive, but by no means basic.
Start building today. It’s free.
