Watson Understands Quebecois and English from Down Under

You can now serve more of your customers across the globe using the new Watson Speech to Text dialects, Canadian French, and Australian English. You can leverage these models for both narrowband audio, usually found in call centers, and broadband audio, typically found in videos and other media. These models deliver a 30–45% improvement in accuracy. That improvement means you can better understand your customers.

Dialects Matter

We know that when it comes to communication, dialects matter. I’ve learned this the hard way more than once, most vividly on the side of the road deep in Quebec Province. It was the middle of winter, and, in an attempt to pull over and look at the map, my friend drove my car into a snowbank. Luckily, we waved down a roadside assistance truck for help. I had studied French for six years, so I wasn’t worried about traveling in French-speaking Canada. I was wrong. I found myself negotiating a fee and explaining the hitch on my car in French. I explained the situation with my voiture, but when he responded, I had no idea he was saying. I listened, nodding my head, thinking “is this even French?” Luckily, some hand waving, miming, and a much more French-fluent friend got me through the situation. The tow truck yanked my car out of the snow. I paid the driver in American dollars, which he grumbled about but accepted, and we were on our merry way.

I’m sure many of you have had this same experience where you know that you technically speak the same language as your conversation partner, but frankly, you don’t understand what they’re saying. Another example of this kind of miscommunication is simply vocabulary. I have had more than a few laughs with Australian colleagues about words like “whinging” (a word for whining used by Australians). Between individuals, this situation can be funny, but it can lead to serious miscommunication. The same happens when applying AI. When it comes to Speech to Text, lack of dialect support leads to low accuracy and frustrated users. That’s why IBM has trained and released Canadian French and Australian English for Speech to Text.

Improvements for Customer Care

For both of these dialects, we focused on quality improvements for customer care use cases. These models specialize in conversations between agents and customers as well as human-to-machine conversations characteristic of an IVR. The training data we incorporated addresses four domains: telecommunications, financial services, insurance, and retail. During training, we focused on tailoring the acoustic models and expanding the models’ linguistic knowledge. This training led to major accuracy improvements (see table below). This localization means that Watson will better understand your Australian and French-Canadian customers.

Implementation Patterns

There are fewer written than spoken differences between dialects. So, these new models are compatible with Watson Assistant and Watson Discovery. As a result, you can implement Watson Assistant for Voice Interaction (WAVI) using these new models. This solution answers your customers’ questions without human intervention, reducing hold times, and agent workload. You can also integrate Watson Discovery and Speech to Text. Discovery can analyze conversations and help your agents find information. The combination of Watson services will enable you to implement a holistic customer care solution.

Customization Enabled

Canadian French and Australian English are both in GA with acoustic and language model customization. Stay tuned for new neural voices from Watson Text to Speech as well! We plan to release Canadian French and Australian English voices before the end of 2020. In the meantime, try out the new models and send us your feedback. You can create an IBM Cloud account here to get started.

Other Helpful Links

• Here’s a guide on getting started with Watson Speech to Text.
• To learn how to evaluate Speech to Text models, check out this article.