At embedded world, Dolphin Semiconductor will present the WhisperExtractor, a revolutionary product in low-power voice processing. This groundbreaking technology enables continuous listening while drastically reducing energy consumption to a level where it becomes almost negligible, making it an environmentally friendly solution. Inspired by the neuromorphic design of the human brain’s auditory system, WhisperExtractor mimics biological efficiency to achieve unparalleled performance in voice recognition. With an impressive accuracy of 93% on Google’s Speech Commands dataset in English and 97.8% accuracy in Chinese keyword spotting (KWS), it delivers robust and reliable voice processing across languages. By enhancing user experience with seamless, always-on voice recognition and improving accessibility, it contributes to more intuitive and inclusive interactions. Designed for applications like keyword spotting, speaker recognition, and speech-to-intent processing, Whisper Extractor sets a new standard for energy-efficient audio feature extraction in wearables, IoT devices, and smart home products.
The innovation behind WhisperExtractor lies in its use of analog Mel Frequency Cepstral Coefficients (aMFCC) extraction, a transformative approach that replaces energy-intensive digital processing with efficient analog techniques. Traditional methods rely on digital signal processing to perform complex operations like the Discrete Fourier Transform (DFT) and Discrete Cosine Transform (DCT), which are computationally expensive and consume significant power. WhisperExtractor, by contrast, leverages Parseval’s theorem, which ensures energy equivalence between the time and frequency domains, to implement these operations in the analog domain. This shift reduces power consumption by up to 99%, enabling always-on voice monitoring at an operating budget of just a few microWatts on a 32kHz crystal.
The architecture of WhisperExtractor begins with the capture of an analog signal through a microphone. This signal is processed through a series of bandpass filters aligned with the Mel scale, mimicking the way the human ear perceives sound by focusing on low-frequency variations. Instead of relying on digital DFT, the filtered signal undergoes rectification, integration, and pulse counting, efficiently isolating its key features. These steps produce cepstral coefficients, which encapsulate both frequency information and temporal variations, making them ideal for voice recognition and other audio classification tasks.
WhisperExtractor’s compact design integrates analog-to-digital conversion, voice activation detection, and feature extraction into a single IP.
The journey of WhisperExtractor began with a vision to revolutionise audio IP design by addressing the inefficiencies of conventional digital solutions. After rigorous R&D, the company developed a robust, scalable IP now production ready. The first customer tape-out is already done, signalling WhisperExtractor’s readiness for real-world deployment.
For more information, visit: https://dolphin-semiconductor.com/dolphin-design-unveils-an-innovative-ip-for-sound-classification-cutting-down-energy-by-99/
The Electronics Excellence Awards ceremony will take place at embedded world on 12th March at 4.30pm at the Hall 3 Forum.
You’re all welcome to come and see which product will win!
