Neural Decision Processors™

NDP100 - Always-On Speech Recognition

The Syntiant® NDP architecture is built from the ground up to run deep learning algorithms. Embedded with the Syntiant Core 1™, the NDP100 is a low-power, low-latency device that achieves 100x the efficiency levels of traditional CPUs and DSPs, all without requiring a connection to the cloud.

Syntiant-NDP100-front-and-back_edited_ed

The NDP100 runs voice, audio and sensor applications such as keyword spotting, wake word detection, speaker identification, sensor analytics and audio event and environment classification, while consuming only 140μW of power.

Ultra-Low-Power

Micro-watt level active power consumption, while always-on and listening

Compact Size

Extremely small 1.4mmx1.8mm, 12 ball WLBGA package for space-constrained applications

Keyword Speech

A keyword speech interface created from the simultaneous classification of dozens of keywords

Neural Processing

Neural network with over half a million parameters providing large scale parallelism for sensor-based applications

Ease of Use

Multiple, flexible methods to access neural processing engine; full customization of applications and post-processing

Key Features

12 ball WLBGA package for extremely space-constrained environments
Dual PDM microphone input or PCM-over-SPI input
Stereo/mono I2S interface multiplexed with PDM
Direct access to neural network over SPI for sensor applications
Frequency-domain, time-domain & batch input models
16-bit input holding tank with faster than real-time SPI extraction
Embedded Arm Cortex-M0 processor with 112KB SRAM
Integrated clock multiplier and dividers support low-frequency clock source or external clocking

Optimized interrupt and SPI slave interface
Onboard firmware security and authentication

English speech service for keyword training
Software Development Kit (SDK) integrates in any software environment
Training Development Kit (TDK) enables the use of standard frameworks such as TensorFlow for customer-programmed applications
Support for 64 output classifications
Active power consumption of <140 µW while recognizing words