Fintech is evolving fast. Gone are the days of clunky logins and endless passwords. Today, users expect seamless account access, minimal friction and one-click payments.
Clients also require high-quality, responsive customer experience. Money is a deeply personal, sensitive issue for consumers. Even the slightest issue must be resolved quickly, efficiently, and respectfully.
Consumers today can connect, sync their data and make payments to anyone in a matter of seconds. And they expect to receive support and clear up issues just as quickly. Which means long call waiting times and slow responses are out of the question for modern financial service providers.
Read on to see how this fintech company uses Gladia’s AI transcription API to deliver an exceptional customer experience, fast.
Because finances are so important to consumers, fintech companies often field a high volume of support and service calls. In this case, we’re talking millions of hours per year!
Speech-to-text tools are therefore critical. This fintech solution needs to transcribe and translate all these calls automatically, both in real time and after the fact. They can react instantly to issues as they arise, triage the highest-priority issues, and use call data for training and product improvements.
Regardless of volume and language, speech transcription needs to deliver on all fronts to do its job—latency has to be low and accuracy high. Users shouldn’t have to trade one for the other.
Another, very specific issue for this company is speech cadence. Bots tend to speak at varying speeds, with numbers being pronounced in differing rhythms. For example, “one hundred thousand” may be transcribed as “one hundred” and “thousand” separately. Which of course leads to transcription errors.
The goal was a bespoke transcription service that was both highly accurate—especially in dealing with numbers—and fast. Because of its specific product and customers, wrong numbers could result in payments being made out to the wrong account, and amounts being off by a lot. Accurate numbers were paramount.
But they also couldn’t afford to keep customers waiting. Users were calling into financial institutions for quick resolution, and in many cases debt collections relied on this fast service. Real-time latency quickly became the second main objective as they evaluated STT providers.
The company compared Gladia to Google’s Speech-to-Text API and OpenAI’s Whisper. Straight out of the box, they were impressed by Gladia’s combination of speed and accuracy, and decided to move ahead.
Gladia’s CTO, Jonathan Soto, set out to optimize the model using a ‘dual-track’ approach.
“It was a very interesting game to play, and the findings we had during the testing phase have since been implemented in our pre-processing for all Gladia users to enjoy”, says Jonathan.
There are three ways to improve the accuracy of a speech-to-text model:
Jonathan and his team worked two streams to achieve the highest possible accuracy on numericals, while also solving for the stark contrast in speech cadence of different bots.
In post-processing, Jonathan’s team used regular expressions (regex) to match specific text patterns and increase the overall level of accuracy.
Pre-processing is where things get more tricky. A speech-to-text model is split into three parts:
Jonathan’s pre-processing work focused on VAD. More specifically, he focused on silences. When speech is interrupted with silences, transcription models can miss chunks and entire utterances that come towards the end of a passage. This results in severe error rates on numbers.
To solve this, Jonathan used autoregressive prediction to predict whether someone is speaking or not in an audio stream.
Autoregressive prediction bases the speech or non-speech status of the audio on previously processed frames (a frame typically lasts between 10-50 milliseconds). This creates a chain-like structure where a certain token only gets predicted and transcribed if another specific token has appeared previously.
In short, it creates highly accurate transcriptions.
For this particular fintech company, the results were remarkable. By partnering with Gladia, they were able to:
Put together, these figures amount to very responsive and highly effective customer service and support. And in consumer finance, that high-quality service is a true competitive advantage.
Product News
Whether you’re using Gladia’s speech-to-text (STT) API during a free trial or a long-term integration, you care about one thing: getting accurate, reliable transcriptions that work for your product and users.
Case Studies
Fintech is evolving fast. Gone are the days of clunky logins and endless passwords. Today, users expect seamless account access, minimal friction and one-click payments.
Speech-To-Text
Customer service is evolving fast to meet new challenges. Today's clients expect immediate, accurate answers to increasingly specific queries and complaints. Meanwhile, contact centers need to reduce costs, improve efficiency, and maintain compliance…all while delivering exceptional experiences.
