Speech Recognition Test Set

ASR (Automatic Speech Recognition) for many languages is still in development. That is why Le Voice Lab exists, a French association that brings together various institutional players (universities, research laboratories, etc.) and private companies whose common interest is to build an independent ecosystem and common standards to enable France and Europe to remain competitive in the global voice market. It’s not just Europe, around the world there is a substantial quality gap compared to the English speaking world for ASR.

But which ASR works best for business’s customers, and is it good enough for the intended applications? Enterprises are now equipped to easily compare the different ASR engines from global and regional providers. vCon enables a single source of test data to accurately and repeatedly measure speech recognition performance across 100s or 1000s of samples, to gather statistically meaningful performance data.

As it’s a computer file format the vCons can be processed through an Excel sheet or business intelligence application. Businesses can make quantified decisions based on their specific situation. The ‘Rolls Royce’ ASR may be the best with an accuracy range of 94-96%, but the ‘Honda Civic’ ASR is good enough at 92-94% for the intended application. The vCons from the different ASRs can be processed through the business application, and the business results compared, not just word error rates.

A business may receive 95% of their voice calls from 3G mobile networks with a range of dialects. They can build their own vCon test set, run them through the ASRs, and with nothing more than Excel compare the results. It could be that ASR in general is not currently up to the task, this will change given the continued performance improvements, but better to make an informed decision and revisit; than assume ASR is inadequate until the gap with competitors becomes clear and leaves your business struggling to catch up.

vCons democratize an opaque industry, which relies on fear, uncertainty, and doubt to stop the buyer making a quantified decision that is best for their situation.