Embedding AI Voice Control for Automobiles and More

Sensory used its VoiceHub tool to rapidly build a compact wake word and large vocabulary language model tailored for appliance voice control. The TrulyNatural speech recognition and NLU engine were optimized to run efficiently on the Alif Semiconductor Ensemble E3 SoC, which is powered by Arm Cortex-M55 and Arm Ethos-U55. All data stays on device, helping to ensure absolute data privacy for system users . The model can run effectively on the Cortex-M55 alone, but Sensory wanted to increase efficiency by using it with the Ethos-U55. Key optimizations included:

• Partitioning workloads across the Cortex-M55 and Ethos-U55 to maximize throughput. Speech recognition runs primarily on the Ethos-U55, with the Cortex-M55 handling text processing and intent determination.
• Accelerating machine learning functions within the Cortex-M55 with Arm Helium to reduce power consumption for always-on wake word detection.
• Optimizing memory usage and network architectures to minimize bandwidth needs for the Cortex-M55 core.
• Reducing model size using Ethos-U55 native support for INT8 quantization and sparsity. This was critical for fitting into the memory-constrained environment of a microcontroller

These software optimizations allowed the advanced TrulyNatural engine to run smoothly on the Arm Cortex-M55 and Ethos-U55 accelerator on Alif’s Ensemble E3 platform. The solution provides customizable, accurate speech recognition and understanding, while adhering to the tight performance constraints of consumer devices.