WhisperExtractor
Neuromorphic Audio Front-End Enabling Always-On signal processing
µWatts voice & sound classification
“Unleash Your Voice, Unplug your charger”
Only 7 µWatts
Near-zero power consumption
32 kHz RC clock
Always on domain friendly.
Seamless Integration
No software required.
Evaluation kit
Available.
Up to 99% energy saving
while maintaining inference accuracy
Google Speech Command (20 word) - accuracy
Chinese KWS (25 words) - accuracy
Reduce the power consumption from your next VUI application
Smart DMIC
Hearables
Smart speaker
Smart remote
AIoT
FAQs
WhisperExtractor is a disruptive technology addressing a major challenge in the voice-user interaction space – low power consumption.
With the increasing adoption of voice user interfaces in everyday objects, it has become critical to have efficient and low-power solutions for voice processing. Conventional digital signal processing methods to process MFCC require a significant amount of energy to convert analog signals to digital and perform signal processing calculations, making it difficult to implement always-on voice listening without draining the battery.
The WhisperExtractor uses a mixed-signal architecture to efficiently extract the Mel Frequency Cepstrum Coefficients (MFCC) needed for keyword spotting, speaker recognition, and natural language understanding applications. By doing so, it reduces power consumption by up to 99%, enabling always-on listening at a power budget of just µW on a 32kHz crystal. This disruptive technology allows for the constant monitoring of the voice without the need for frequent charging or compromising battery life.
With its mixed-signal architecture, the WhisperExtractor is able to efficiently extract the Mel Frequency Cepstrum Coefficients (MFCC) needed for voice processing like you would usually do on a DSP. The output data of the WhisperExtractor is fully compatible with modern machine learning models and accelerators, including CNNs and RNNs.
This means that it is a versatile solution that can be easily integrated into a wide range of applications, from smart home devices to wearables and beyond.
Overall, the WhisperExtractor is a truly disruptive technology that is changing the game in the voice user interaction space. By offering low-power voice processing and always-on listening, it opens up a world of possibilities for developers and manufacturers looking to create innovative new products. So if you’re looking to revolutionize the voice user experience and stay ahead of the competition, the WhisperExtractor is the solution you’ve been waiting for.
The WhisperExtractor is designed to be easy to integrate into your SoC. It comes as a mixed IP that is delivered as GDSII and RTL, which means it’s ready to be integrated into your design without any additional effort required. This plug-and-play solution makes it easy for you to start using the WhisperExtractor quickly and without any specialized expertise.
Once integrated, the WhisperExtractor will output the MFCC coefficient without the need for any programming or additional know-how. This means that you can start using the extracted features right away without having to worry about the details of how they are calculated.
However, we understand that different applications have different requirements, and that’s why we’ve built great flexibility into the WhisperExtractor IP. This allows you to adjust the parameters to suit your needs, giving you the ability to optimize the feature extraction process for your specific use case. With this flexibility, you can fine-tune the WhisperExtractor to deliver the best results for your particular application, making it a highly adaptable solution for a wide range of use cases.
To help you evaluate the whisperextractor IP, we have developed a Matlab or Python model that can be integrated into your DNN flow. This model will allow you to test the IP’s performance in your specific application and environment. Playing with this model will show you how easy it is to use the wisperextractor compare to conventional solutions and how much value it adds.
Our Matlab and Python model provides a powerful tool for evaluating the whisperextractor IP. You can use it to experiment with different settings, measure performance, and develop and refine your DNN architecture, leading to faster time-to-market and improved results in your specific application.
If you would like to run your own evaluation or have any questions about the whisperextractor IP, please don’t hesitate to contact us. Our team is always available to provide support and guidance as needed.