Behavox Research on Continual Learning Featured in ICASSP

MONTREAL – May 16, 2022 – Behavox, which provides a suite of security products that help compliance, HR, and security teams protect their company and colleagues from bad actors, today announced that its academic research paper on artificial intelligence “Continual learning using lattice-free MMI for speech recognition” has been accepted by the International Conference on Acoustics, Speech, and Signal Processing (ICASSP). ICASSP is the world’s largest and most comprehensive technical conference focused on signal processing and its applications. Behavox presented this paper in Singapore for IEEE ICASSP 2022 on May 9, 2022.

“Sometimes scientists do research for science’s sake and society is unlikely to see benefits in the near future. In contrast, we pushed science to a new frontier to find a solution to a current problem,” said Behavox CTO Joseph Benjamin. “We are honored to have received recognition from such a prestigious conference and are very proud of our team and the work they did.”

The paper, written by Behavox’s Machine Learning team, addresses challenges of continuously improving speech recognition systems in privacy-preserving settings. This research directly links to Behavox’s R&D approach for providing customers with the best transcription quality in the financial market. In FinTech, the language and the data that Behavox systems have to reliably process rapidly change over time, with the rise of cryptocurrencies and video communications being but two examples.

Consequently, machine learning models must be frequently updated on relevant data. To make such updates possible without undermining data privacy, the model should be able to gradually improve inside secure client environments without access to historical data and data from other customers. This is a challenging problem in machine learning due to catastrophic forgetting, where after improving on a specific type of data, the model starts to perform badly on previously learned data.

The Behavox ICASSP paper contributes to the field of continual learning for speech recognition. Authors propose a novel algorithm that extends and refines the commonly used Learning Without Forgetting (LWF) technique, which relies on a regularization term and forces neural networks to reduce forgetting. More specifically, the authors designed a novel sequence level neural network training loss that can be used instead of conventional point-wise LWF. They demonstrated a significant reduction of forgetting when the neural network is fine-tuned sequentially on various accents and speech styles.

Behavox has found that the use of voice platforms has increased 94 percent year-over-year and that on average almost a third of all workplace communications take place via phone or video conference calls.

This technique has already been successfully applied to its Danish system, where the collaborative work of Behavox linguists, the machine learning team, and a customer allowed Behavox to build its first Danish transcription system that reliably works on banking data.

Read this article in Swedish or Danish:

Swedish: https://www.businesswire.com/news/home/20220513005573/sv
Danish: https://www.businesswire.com/news/home/20220513005572/da