Neural network-based speech recognition technology for IoT devices

Leveraging CEVA’s vast expertize in low power voice and audio processing, WhisPro is an always-listening multi-trigger phrase technology allowing users of new smartphones, smart speakers, Bluetooth earbuds and other voice-enabled devices to interact with cloud-based voice assistant services, such as Amazon Alexa, Google Assistant, Baidu DuerOS and others. CEVA provides its customers with a trigger phrase customisation service to enable voice control for a range of use cases and end markets, including automotive, smart home, and enterprise.

WhisPro operates in tandem with CEVA’s ClearVox front-end voice processing software technology and offers its customers a robust CEVA DSP-based hardware and software solution for speech recognition. This holistic integration of voice pre-processing and neural network algorithms delivers a high recognition rate of more than 95%, while consuming minimal computing resources and power. The technology shows resilience even in noisy environments in both near field and far field use-cases. WhisPro technology operates locally on the edge device, without the need for cloud back up thereby maintaining user privacy and providing instantaneous response with minimal latency.

Erez Bar-Niv, Chief Technology Officer at CEVA, commented: “Voice has rapidly become the human interface of choice for the intelligent device revolution, yet is one of the most complex technologies to implement. In designing WhisPro, we have capitalised on our low power DSP cores and our in-house noise reduction and echo cancellation algorithms to provide our customers with a complete speech recognition technology that is cost-efficient, highly-accurate and noise-immune. It supports edge and cloud-based voice assistant services and also enables customers to implement their own proprietary trigger phrases to control their devices.”

Based on speech AI technology and optimised specifically for edge devices, WhisPro’s scalable Recurrent Neural Network technology is capable of recognising multiple trigger phrases, simultaneously. The WhisPro SDKit provides customers with guidelines and tools to integrate and test the WhisPro technology in their system. WhisPro supports multiple pronunciations of keywords and multiple languages, per customer request, with English available now and Mandarin Chinese available shortly.

WhisPro is available for licensing today, exclusively for CEVA-TeakLite-4, CEVA-X2 and CEVA-BX DSP customers, and is compliant with the main tier-1 voice assistant services.