Detecting sound events is crucial in a variety of industries, from automotive, security to healthcare and others. However, solutions are often required to run on tiny embedded platforms, that’s why we are excited to introduce Low-Power Sound Identification – a machine learning technology that can detect sound events using minimal power.
Our low-power SoundID solution can run on DSP chips, such as the Cadence HIFI5, which are designed for low power consumption and high-performance audio processing. These chips allow us to provide real-time analysis of audio signals and accurately identify sound events such as sirens, coughing, and crying babies, and many more.
We use deep learning algorithms to build small models that can be around 100kBytes or less. These small models enable us to achieve high accuracy while maintaining low power consumption. We use TensorFlow Lite Micro, a lightweight machine learning framework, to efficiently implement these models on low-power devices. This enables us to provide customized solutions for each application, tailored to their specific needs.
For applications that require even greater accuracy or need to operate in very diverse acoustic background conditions, we also offer application processor solutions and cloud solutions for SoundID. Here we provide exceptional accuracy and performance, making them ideal for use in complex sound detection applications. These more powerful systems can also be used as a revalidation of the low-power SoundID detector, ensuring the highest level of accuracy and reliability. In many cases, especially in predictable acoustic environments, low-power SoundID on DSPs can already be accurate enough without the need of a revalidation stage.
This new technology has a wide range of applications, from smart homes and security systems to healthcare and industrial automation and can be ported to many different platforms. We are excited to bring this technology to the market and to see the impact it will have on our daily lives. Contact our team to learn more.