The accuracy of speech-to-text models is improving all the time, with platforms like FinTalkrTM now boasting 95% accuracy over a 2-hour technical discussion between a financial adviser and their client.
The accuracy of speech recognition systems is usually based on Word Error Rate (“WER”). Performance is complicated by many factors, such as accents, pronunciation, and background noise. Error rates also increase as vocabulary size grows, and it is not uncommon for 100,000 words to have error rates up to 45%.
We’ve come a long way since the 1950s, when initial attempts to recognise speech covered a few hundred words and relied on pauses after each word. By the mid-1980s, machines were able to handle 20,000 words using a statistical approach called the Hidden Markov Model (“HMM”).
Today, despite your average in-car system only using a fixed set of commands to perform basic tasks such as initiating phone calls, most sophisticated speech recognition systems are now using a deep learning method called Long Short-Term Memory (“LSTM”). This has increased vocabulary size and decreased error rates significantly.
FinTalkrTM is a conversational-driven AI platform specifically designed for financial advisers to transcribe your client meetings in real-time and produce a detailed file note by advice topic that meets your standards. The latest version of FinTalkrTM includes a superior speech-to-text model that even picks up numbers to the cent.
–
If you’d like to hear more about FinTalkr or see it in action, please don’t hesitate to contact the team on info@fintalkr.com