"DSP Group offers compelling hardware for many low-power applications, so we're excited to collaborate to offer machine learning software to help enable developers create products that wouldn't be possible otherwise," says Peter Warden, Staff Research Engineer at Google.
TensorFlow Lite for Microcontrollers is an extension of Google's TensorFlow Lite that addresses the need to run ML on memory-constrained devices with only kilobytes of memory. It comes with a specific set of optimized operations to allow the execution of ML models for applications such as wake-word detection, sound detection, and image wake-up. In addition, developers can add their own ML algorithms. To further enhance efficiency, the DBMD7 has a floating point unit (FPU) for each of its high-frequency cores so TFL code can be executed optimally using either floating-point or fixed-point functions.
"By adding TensorFlow Lite for Microcontrollers to our SDK for the DBMD7 family of AI and DSP processors we allow designers to leverage Google’s open framework and tools to execute advanced audio and voice ML inference at the edge without compromising on cost or performance," says Yosi Brosh, CVP and head of SmartVoice Product Line at DSP Group. "At the same time, they can take advantage of a platform that scales from two to eight microphones for far-field voice user interfaces (VUIs) and voice communications, which are supported by advanced audio processing algorithms based on our 30 years of experience and technical support in this space."
The announcement of the porting of TensorFlow Lite for Microcontrollers comes just days after DSP Group announced that its SmartVoice solutions, which include the DBMD7, reached the 100 million units shipped milestone. The porting further expands SmartVoice’s base of possible applications, which already ranges from AI/ML at the edge, to remote controls, smart speakers, hearables, tablets, laptops, smartphones, and security systems.