New SoC for Speech and Noise Recognition With Lowest Power Consumption
Katana + Video Processing
Ultra-low-power edge solutions require localised intelligence that relies on a unique combination of power-efficient hardware and advanced AI algorithms, while being easily adopted for new applications. The DBM10L adds video processing to the functionality of Katana, another already existing SYNAPTICS product.
Example Applications Created to Demonstrate Effectiveness of Katana and DBM10
SYNAPTICS developed example applications that show how effective both Katana and DBM10L platforms and associated tools are at enabling training models for specific use cases for the IoT. They are well suited to help accelerate the adoption of tinyML for the microwatt era of smart and flexible battery-powered devices across a wide range of industries.
SYNAPTICS DBM10L is an ultra-low-power, small-form-factor, cost-effective artificial intelligence (AI) and machine learning (ML) SoC based on a digital signal processor (DSP) and neural network (NN) engine, both optimised for voice and sensor processing. It is suitable for battery-operated devices such as smartphones, tablets, wearables, and hearables, including true wireless stereo (TWS) headsets, as well as smart home devices such as remote controls.
The DBM10L can enable AI/ML, voice and sensor fusion functions that include voice trigger (VT), voice authentication (VA), voice command (VC), noise reduction (NR), acoustic echo cancellation (AEC), sound event detection (SED), proximity and gesture detection, sensor data processing and equalisation.
The DBM10L’s NN engine comprises DSP Group’s nNet Lite NN processor, a standalone hardware engine that is designed to accelerate the execution of NN inferences. nNet Lite gives the DBM10L its ML capability and is optimised for maximum efficiency to ensure ultra-low power consumption for small to medium-sized NNs.
The DBM10L is supported by embedded memory, as well as serial and audio interfaces for communication with other devices in the system, such as an application processor (AP), codecs, microphones, and sensors.
LOW POWER
- A highly-compact form factor: ~4 mm²
- Ultra-low-power inference consume ~500 µW (typical) for voice NN algorithms
- Runs Hello Edge 30-word detection model @ 1 MHz (125 MHz available)
- Allows porting of large models (10s of megabytes) without significant accuracy loss using model optimisation and compression
Use cases
The platform can be used to process the following sensor inputs: sound, voice, temperature, water, pressure, current, motion, position, etc. Applications are remote controls, voice wake up systems in AOV applications, security functions, sensor input co-processing,
SW Development
- The DBM10L comes with a SW framework, which has interfaces to the market standard AI and ML development tools. These models are then optimised with a compiler for the DBM10L Neural network engine.
Your Contact Person
For more information, please contact Achim Stahl.
