We’re now extremely accustomed to ‘social proof’ as a concept. When looking at whether or not to engage or buy from a company, we often sift through testimonials, read Trustpilot ratings, or stumble across hilarious customer reviews like this one for the 571B Banana Slicer:
As a brand or business, this feedback is great, but obviously is only a tiny window into the minds of customers. Reviews tend to skew either extremely positive or negative, since everyone in the middle probably did not have an emotional enough reaction to leave a review in the first place. Reviews are only left by people choosing to take the time to write one.
But what about what people are saying (literally)? Contained within video and audio, or recorded conversations, there might be a blend of different sentiments.
"The build quality of the banana slicer is incredible, but it was delivered in red, not yellow, but it arrived just 20 minutes after I ordered it, which was super valuable as I had a lot of bananas to slice."
A three-star rating doesn’t capture this nuance, but a 10-minute recorded customer call would.
Sounds great in principle, but how are you supposed to capture this at scale? If you are a call center fielding thousands of customer calls a day, or week, how can you even begin to make sense of all that nuanced sentiment?
Well, enter Sentiment Analysis.
Sentiment Analysis: How people really feel, at scale.
Sentiment gives businesses the ability to track and analyze sentiment for any transcribed media via an API call.
Our Sentiment Analysis uses machine learning to look at complete sentences or segments of an audio transcription and ascribes a positive, neutral, or negative sentiment classification, along with a confidence score between 0 and 1. For example, if a sentence was given a ‘positive 0.92’ classification, that would be a very strong indication that the model predicted the segment to be positive.
This isn’t the only piece of insight you get with Speechmatics. For any audio transcript, you get several pieces of analysis:
Segment analysis: Each transcript is split into segments, with the sentiment of each of these segments highlighted. A conversation might go from negative, to neutral, to positive, and back again. This would allow you to see that shift quickly.
Speaker sentiment: If you’re using Speaker Diarization where different voices are used to identify speakers, you can also access the overall sentiment of a given speaker for the interaction.
Channel sentiment: If you’re using multi-channel media for speaker separation, you can receive the overall sentiment for a given channel for the interaction.
Overall sentiment: If you’re looking for a high-level view of an interaction, this will give you an overall sense of the sentiment.
Confidence scores: Language is nuanced, so we also provide you with some important context for any sentiment, indicating a level of confidence in our classification (between 0 and 1).
To showcase this, we’ve run a news segment from Bloomberg through our Sentiment Analysis:
You can see clearly that the sentiment shifts, even within one speaker’s single answer. And, although the above seems largely positive, it should perhaps not come as a surprise given it’s a news segment, that the overall sentiment of this conversation was negative:
Specific Use Cases
Any organization looking to explore how their customers are really feeling about their service will find this useful, but here are a couple of specific use cases...
Contact Centres and CCaaS
Recorded customer calls can both be a fantastic source of insight and resource used for training. In terms of insight, businesses can analyze the source of both negative and positive sentiment (was it a delivery issue, or a customer service issue?). For training, agents can be trained on calls where negative sentiment was successfully mitigated and handled by an agent, and lessons can also be learnt from where calls take a turn for the worse.
Across broadcast news, social media, podcasts, and other video and audio platforms, there’s plenty of different sentiment going around. Using speech-to-text in combination with a Custom Dictionary (to pick up those more technical or personal words) and Sentiment Analysis can ensure that brands keep informed not only with what people are writing about them, but what they are saying and what their sentiment is too.
A growing suite of Speech Understanding capabilities
Sentiment follows hot on the heels of Summarization and will be a part of a growing list of speech understanding capabilities we're launching here. Next up is Topics, so watch this space for that one.
We're building these because we're now at a point where transcription is only the starting point for a lot of valuable activities. We're on a mission to keep adding features and functionality that allow our customers to make the most of their voice data, and drive value in new and interesting ways.
We hope it goes without saying that capabilities like Sentiment only become powerful when the transcription itself is accurate. When accuracy drops, so does the accuracy of subsequent Sentiment Analysis, and any further downstream activities.
With Sentiment powered by unmatched transcription accuracy, you'll never miss a thing when it comes to how customers really feel. Social proof will never be the same again.
To explore in detail how to implement Sentiment, please head to our docs.