“OK Google, set the alarm for 6:50.” “Alexa, what will the weather be like tomorrow?” Voice assistants are becoming ever more common. The Algorithms and Data team, experts in language and speech processing technologies, has developed a voice assistant that runs locally, without an internet connection: Macarena.
The acquisition of SNIPS by SONOS slowed down projects based on that technology, whose strength was that it worked in French and English and could be adapted to several use cases.
Macarena is based on speech-to-text, which can be developed with 2 methods: deep learning, and a hybrid method combining deep learning with mathematical voice analysis methods
Macarena can be used in “WakeWord” mode (activated by a keyword, like “Ok Google”) or in Streaming mode (the assistant responds to requests without being activated first).
Here we present the technology in several languages and on a use case from our partner La Banque Postale.
The 2 existing languages (French and English), which can be used interchangeably, are handled with NLU (natural language understanding) for smooth exchanges between the user and Macarena. New languages can easily be added with a suitable dataset.
The result is a lightweight “offline” voice assistant whose data never leaves the device, and which understands (and answers in) several languages. Through various demonstrators, we show functional coverage equivalent to SNIPS.
There are other speech-to-text solutions on the market, but Macarena is now one to reckon with! As well as being reliable and robust, it is the only voice assistant technology that is fully adaptable, both in the languages recognized and in the technical terms it can understand.
Like the first voice assistant developed by CATIE, Macarena can be used to make bank transfers, but also to control machine tools, interact with robots and more: in short, in any industry, and even embedded in factories.
Our experts are already working on a beta version using the hybrid method, which will improve performance, handle accents better, understand incomplete or poorly structured sentences, and cope with noise or multiple voices.
Macarena could then be used to assist call center agents or callbots, for example.
While Google and Amazon offer finished voice assistant products, we provide a technology that can be adapted and transferred to any professional or industrial sector.
Embedded artificial intelligence is one of CATIE’s strategic themes, and Macarena illustrates it perfectly, in line with the regional initiative on the industry of the future.
To find out more, contact us !