Researchers at the Japan Advanced Institute of Science and Technology (JAIST) have developed a new system that mimics the brain’s activities to make the speech of virtual assistants more natural. Algorithms are able to capture the complexity of a conversation and transform it into more coherent signals, as happens between human beings.
They managed to create an artificial intelligence (AI) device that emulates this brain function, using an advanced matching search system to obtain sparse representations between speech signals that have as few significant coefficients as possible.
“In humans, the auditory periphery converts the data contained in incoming speech signals into patterns of neural activity that the brain is able to identify. We used psychoacoustic principles such as an equivalent bandwidth scaling and masking effects to ensure that the auditory representations were more similar to our own.”
Want to stay on top of the best tech news of the day? Access and subscribe to our new youtube channel, Kenyannews News. Every day a summary of the main news from the tech world for you!
copying the brain
To test the model’s ability to understand voice commands and generate a more understandable and natural response, the scientists compared the quality of signal reconstruction and the structures of auditory representations with other methods already used in speech recognition systems.
They reconstructed 630 samples of voices pronounced by several different people, resynthesizing the signals with classification models within scoring scales known as PEMO-Q and PESQ — objective measures that assess the quality of the emitted sound.
“As in the human brain, the effectiveness of an auditory representation can be evaluated in three distinct aspects: the quality of the resynthesized speech signals, the number of non-zero elements and the ability to represent perceptual structures,” adds Unoki.
natural voice
By using the algorithms to emulate brain patterns, the researchers found that the resynthesized signals are comparable to the original ones, making digital communication much closer to a natural conversation seen in the interaction between two humans.
Using a pattern-matching experiment to determine whether auditory representations could be matched to spoken sentences, the scientists found that the system was able to capture voice structures more accurately, making dialogue more coherent and less robotic.
“The model developed in our study can go a long way in conveying human qualities, such as different pronunciations and accents, to our voice assistants, making our digital interactions not only more convenient, but also psychologically more satisfying,” concludes Professor Masashi Unoki.
