Introducing Flow – passing the Turing test with speech

A new era in Conversational AI

Today marks an exciting new chapter for Speechmatics. 

We’re bringing all the expertise that has allowed us to lead the ASR market for a decade into the world of Conversational AI and AI assistants. 

Today we are introducing the world to Flow. 

Flow enables any company to build incredible speech interaction into their products.

It has a huge vocabulary, and is consistently accurate even at incredibly low latency (so no waiting for responses). It also includes highly accurate speaker identification which enables game-changing interactions that include multiple people or in noisy environments. 

The range of use cases for this is vast. It includes generalist AI assistants, but can also be used for any AI agent or product that would benefit from people being able to speak to it.

Why are Speechmatics doing this?

Does it change our mission to understand every voice?

For us, the answer is simple – our mission remains, but the scope of our product just got slightly bigger.

The intuitive need for speech interactions 

Why do we need speech interactions? 

There’s a simple, and strong, reason to want to add speech interactions to our current roster of ways to use technology – it is deeply intuitive.

We’ve been using our voices for 100,000 years – far longer than any tool, let alone a keyboard and mouse.

It is the default way that we, as humans, communicate with each other.

In a world where technology is becoming increasingly powerful, technology that can ‘meet us where we are’ will ultimately be far more valuable than technology that requires a degree in computer science to use.

Technology we can work alongside and collaborate with to perform tasks, solve problems, complete tasks, all without lifting a finger (literally) will transform the way in which we work and play.

Addressing accessibility challenges

For many, our current methods of using technology will seem straightforward to use. A (touch) screen displaying information and images, and a mouse and keyboard to complete tasks.

Simple enough.

Except, for millions, it is not.

Worldwide, there are millions of people for whom this option presents significant barriers:  

• A World Economic Forum study in 2017 revealed that almost a quarter of respondents did not know how to operate a computer.

• 3 in 10 Americans struggle to use the internet.

• 10% of the world’s population has dyslexia (that’s 780 million people).

• 40 million people are blind, with a further 250 million having visual impairment issues.

• 7% of working adults have dexterity issues and will struggle to use a keyboard and mouse.

As well as the exciting opportunity to give people a new way to interact with technology, there is also the motivation of creating truly inclusive technology that all but eradicates the traditional barriers of adoption.

The current landscape

There may be some thinking that all the key AI players are implementing speech as the core interface to their intelligence stack.

There may be many die-hard users of Siri, or Google Assistant, or have tried ChatGPT since their introduction of GPT-4o. 

There may be those thinking that speech is solved. 

They would be wrong.

An easy way to show this is to imagine the Turing test.

First proposed in 1950, it is a test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.

Another way of thinking about this is that if a machine can ‘fool’ someone into thinking it is another person, the test has been passed.

For a few years now, it can be argued that we’ve reached this point with text interactions. It is very difficult to speak via chat with an LLM and not forget that there is no person on the other side.

But imagine now trying to pass this test using speech. Are we at the stage where someone would be able to conduct a conversation (say on a phone) with an AI, and come away believing they had been speaking to a person?   We believe the answer to this is no.

Building future-ready technology

Live human conversation is messy.

Picture yourself sat in a bustling coffee shop, texting a group of friends.

The grammar of the interaction is well understood. Someone else in the group says a sentence or two and sends out a message, you read it and decide to respond. You write a few words, send it, and wait. Each sent message usually represents a complete thought, with the person deciding when to send it. The conversation flows slowly and steadily.

Now imagine sitting in that same coffee shop with the same friends, having an animated conversation.

It is far faster, more dynamic, and full of subtle cues to make the interaction function.

Face to face conversations are full of emphasis, intonation, accents, slang, dialects, interruptions, cross talk, pauses, stammers, stutters, noises, laughter, sarcasm, and more. 

This doesn’t even include the fact that everyone there recognizes each other’s voices (in theory, you could all close your eyes and continue the conversation).

It also doesn’t include the fact that those at the table can distinguish between those a part of the conversation and background chatter from the rest of the coffee shop. 

These characteristics, while deeply intuitive and understood by people after a lifetime of social interactions, are really tough challenges for technology to understand and overcome.

Imagine opening Siri, Google Assistant, or ChatGPT in that coffee shop and making it a part of the conversation.

Do you think it would blend right in and be one of the groups?

Or do you think it would fall over, its basic function broken by the number of curveballs being thrown its way? 

We still have a way to go. 

What’s also important to note is that the ‘speech-in’ part of any interaction with AI is the most important link in the chain.