Xiaomi has developed an ultra-fast voice recognition model and made it OpenSource
Xiaomi has developed a voice recognition module called MiDashengLM-7B. By using neural networks instead of fixed algorithms, the company achieved the fastest voice recognition performance in 22 synthetic tests. This makes it possible to build user platforms that work with almost no delay. The model can be used in smartphones, smart home systems, cars, etc.
MiDashengLM-7B analyses audio on the fly, separating ambient sounds or music. Xioami is already actively applying this voice model in practice in its products, for example, the YU7 car constantly analyses sound and can detect the sound of scratching or breaking glass, which allows you to turn on the alarm even when there is no impact that would be detected by the motion sensor.
Xiaomi has published the source code for the advanced voice under the Apache License 2.0, as well as detailed documentation on the training and implementation of the technology. The model can serve as a basis for developers and academic researchers seeking to create open voice systems without dependence on closed ecosystems.
Chinese companies are not known for working on open source projects. By making the language model open, Xioami attracts more developers, which will help this product compete on an equal footing with analogues from large technology corporations. Experience has shown that large and complex software products, such as an operating system or browser, develop faster and become more competitive if they are developed by an open community of programmers rather than a single company.
Source: xiaomitime.com
