Great Improvements for Brazilian Portuguese and Canadian French ASR

There’s something hugely satisfying about overtaking your competitors after putting in the hard work. It’s even more of a win when you find yourself way out in front, as we have with our latest uplift across two of our most often used languages French and Portuguese. In particular, looking at the significant improvements to these models and what they mean for Brazilian Portuguese and Canadian French ASR.

Using a wide range of test sets, we’ve challenged both our Enhanced and Standard models on two fronts. First, we see how both models test against our last update from earlier this year, then it’s up against rivals from Big Tech, like Google, and Microsoft.

Huge Uplift for Canadian French

Let’s begin with our hugely impressive uplift concerning Canadian French. In Quebec alone, it’s believed over 7 million people speak a variety of Canadian French, making up 22% of the Canadian population. A further 2 million speak it as a second language. Using a 6-hour dataset from Journal des débats de l'Assemblée nationale, we saw an average relative reduction in Word Error Rate (WER) of 18% for our Standard model and 13% for our Enhanced.

What’s more, the huge leaps we’ve made with our Standard model mean we’re now ahead of all our competitors, as illustrated in the table of results below. (We’ve displayed these results in terms of accuracy of speech recognition for Canadian French, which is simply 100 – WER).

Continued Improvement for Brazilian Portuguese

We’ve been equally pleased with our latest round of results for Brazilian Portuguese. With over 200 million Brazilian Portuguese speakers, around the world, it’s essential that they’re serviced by great ASR. Speechmatics can proudly claim that we deliver great ASR for Brazilian Portuguese.

On two internal test sets we see an average relative reduction in WER of 5% for our Standard model, and 3% for our Enhanced model, compared to our last quarterly release. While we’d love to say we’ve again overtaken our competitors, we can’t. Because we’ve always been ahead of them when it comes to Brazilian Portuguese! These latest results increase our market lead, smashing the 80% barrier, where other much larger names hover around 70%.

The Benefits of Self-Supervised Learning

As we continue to use self-supervised learning, we’ll continue to see the benefits. With a number of new languages on the way to add to our already impressive list of 34, we’ll do our best to improve on every one of them continuously. After all, our mission is to understand every voice.

Voice by voice, we’re getting there.

John Hughes, Accuracy Team Lead