How Yandex Cloud developed a polyglot neural network
The new machine learning model can recognize up to ten languages at once, including English, French, Danish, Finnish, and Turkish.
We focus on providing businesses access to the dialog scenarios used to develop voice assistants like Alice, automate customer service, run surveys, and provide quality control. No matter the feature, speech technology demands huge data sets, highly qualified professionals, and hundreds of hours of training. Yandex Cloud is constantly working on improving Yandex SpeechKit, our speech synthesis and recognition service. We’re making speech more natural, expanding our capacity for recognition, and creating unique assistant voices.
The new machine learning model can recognize more than ten languages at the same time. Using the new neural network, companies are creating multilingual voice assistances and robots for high-volume call centers faster and easier than ever before.
The polyglot neural network goes beyond popular languages like English and French to also cover rarer options like Danish, Finnish, and Turkish. The model recognizes speech regardless of the topic and format: short and long phrases, names, addresses, dates, and numbers. It even recognizes languages automatically, switching between them at a moment’s notice. With that said, giving it hints improves recognition quality.
The new multilingual model is built on transformer architecture. As transformers process incoming words in parallel and independently of each other, the learning process takes far less time. The tens of terabytes of data in the processional data sets and arrays pulled from Yandex services and apps and used to train the neural network ensure fluid, grammatically correct text in the target language.
New multilingual recognition is already available for Yandex SpeechKit users and configured as usual via the API.