I really wanted to explain about some of the common misconceptions when using this technology within the industry. Accuracy is an obvious measure for speech recognition and the use of word error rate (WER) is considered to be the way to measure accuracy but this can be misleading. I wanted to challenge this academic understanding and explore, from Speechmatics’ experiences, what accuracy really means when applying it in real-world applications. In the Webinar, I talked about how an ‘accurate’ speech recognition system can mean vastly different things from company to company and their differing use cases. It was really great to demonstrate integrations that we have with our partners TranscribeMe and Red Box Recorders to show the broad application of speech recognition.
These demonstrations highlighted some of the common challenges that we face in the industry and showing that the term ‘accuracy’ has far more complexity than simply WER.
The Webinar was hugely successful with lots of attendees eager to learn more about ASR, our language coverage and indeed how we are striving to become more ‘accurate’ across many applications.
I spoke at great length about how we at Speechmatics are not only continuously pushing our WER down but are also developing cutting-edge features and efficiencies to improve our accuracy for multiple real-world applications, from readability and punctuation to speaker diarisation and customised dictionaries. For those that missed the Webinar, or for those that would like to watch it again at your leisure, we recorded the Webinar in full for you. The recording includes the content from my discussion, as well as the Q&As from the registration submissions. You can view the recording of my Webinar below. Interested to hear more about ASR accuracy and how we’re working to improve the industry benchmark for real-world applications? Speechmatics will be addressing the topic of accuracy over the coming months, so stay tuned on our social channels and blog to see what is heading your way next. Ian Firth, Speechmatics