Did you know there are over 800 million videos on YouTube, with over 30,000 hours uploaded every single hour? These include channels dedicated to:
Squashing things in a hydraulic press
Smashing things in ultra-slow motion
Popping pimples (do not look this up if you are squeamish. I mean it.)
Asking a mortician questions
Eating food very close to a microphone
Each to their own...
Now imagine trying to find out what every YouTube video was about – not just the whole video, but segments within it. Now imagine doing that across all social media channels. Now imagine doing that across every single customer call that's ever been recorded (69% of businesses record all of them). Imagine if you wanted to find all conversations said on a topic like 'politics'.
It boggles the mind 🤯
Unlocking Insights
The above examples are perhaps edge cases, but for businesses looking to make sense of media in this way, there's a real challenge here, with plenty of value to be had if they can successfully achieve it. For them, those topics may be things like 'cost of goods', 'customer service', or 'social media'.
Find out what large groups of customers are saying about those things, and you'll grow faster than the popularity of the hydraulic press channel (8.78 million subscribers and counting).
If you guessed that Large Language Models, Natural Language Processing, and Speechmatics might be part of the answer, then ten points for you.
Speechmatics has launched Topics, allowing our customers to parse transcripts and analyze them for the topics being spoken about.
Read our docs, or continue this blog to learn about how it works and some great use cases.
So, how does this all work?
Analysis of Speech at Scale
Topics (sometimes called Topic Analysis or Topic Classification) uses Large Language Models' to quickly identify custom topics in your content and accurately discover what's being spoken about. What's clever about this is that it's doing far more than scraping a document for specific words and returning the sentence that contains them.
It can see a sentence like this:
"It's really important to pick a mortgage with a fixed interest rate if you want to avoid increased payments due to rising interest rates."
And know that it's about finance, without any of those words appearing in the sentence itself. This obviously has some major advantages, since we often discuss a topic without mentioning it by name. This can be done at a speed and scale that humans simply cannot match.
In our case, that written text can be transcripts of media, customer conversations, social media videos, podcasts, internal meeting recordings, and more. This opens up the potential value of Topics to a much wider set of use cases than if it were only restricted to written content.
Here's an example of this in action, where we have transcribed a news segment taken from BBC News:
Here we can see the segments identified and each section is given a topic that it covers. These topics were entered before Topics ran – we could have easily chosen to see other topics like 'Finance', 'Hydraulic Pressing', or whatever else we were interested in. Don't forget that timestamps are provided so you can easily see exactly where in the conversation this topic was covered.
Topics in Speechmatics
Our Topics API is extremely flexible – users can enter the shortlist of topics of their choice and have transcripts scanned and analyzed for those topics.
This is because we harness LLMs' inherent knowledge to quickly identify a wide range of topics to enable customer-specified topics, not just a set of pre-defined ones. This avoids the need to gather labeled speech data and train a bespoke machine-learning model. Importantly, this also creates an enormous amount of flexibility and therefore valuable insight, since you're not forced to simply scan a document for discussions of 'food' or 'economics' – whatever you want to sift for, you can.
When you request Topics via our API, you will get:
The list of topics that were detected
How many times they were detected
At what time(s) the topics were detected in the transcript
The related transcript text which discussed the topic
As with all of our features, Topics is available via our unified API, so you can have an analysis returned alongside your transcript with a single API call 💪.
Use Cases
Contact Center as a Service (CCaaS)
Customer conversations are rarely about one thing in particular. A single phone call might discuss everything from delivery times, returns, customer service, price, warranties, and the weather, to release dates of future products. Topics can therefore make conversations far easier to parse and categorize, with the transcript still being readable to discover the precise wording the customer used on a particular topic.
This can be used to improve processes and customer experience, as well as train individual agents and be a great source of feedback on aspects of a business' offering.
Media Monitoring
Basic Media Monitoring can identify specific words and pull out the conversation that contains a specific keyword.
Great Media Monitoring would identify the topics that surround conversations around those words, which could then go on to inform brand workstreams and advertisement targeting. For example, a shoe company might be looking at their make of sneakers, but also want to know that people tend to bring it up when talking about 'yoga', 'hiking', and 'lifestyle'.
