Voiceprint, enabling authentication and personalization on the edge
Updated: Sep 22
Just like a fingerprint, our voice is unique to each of us. With modern machine learning and compute optimization, voiceprint recognition can now be deployed on wearable and edge devices, even under demanding power efficiency constraints.
For a lot small devices, fitting a user facing camera or fingerprint sensor often doesn't make sense. However, having a microphone and listening to our voice is an effective way for the device to know us. Being able to identify the user(s) opens up whole new space of use cases, from authentication to personalization by user profile.
Banking apps on smartphone have been using voiceprint along with facial recognition to authenticate transactions for some time. Replicating this on small devices, however, is considerably more challenging due to resource constraints. For voiceprint to be viable on a 64MHz MCU, for example — resource footprint and inference efficiency must be addressed throughout the pipeline, from initial model design to final deployment. Lower resource requirements unlock exponentially more possibilities:
Ultra low power audio devices, earbud, hearing aid
Home automation/security, such as smart doorbell
Wearable, health devices
The MLSP implementation of voiceprint combines modern machine learning and conventional signal processing techniques (for inference optimization). The result are:
Tiny resource footprint, 20MHz on ARM MCU and 6MHz on Hifi DSP, 100kB RAM.
Sentence independent, any 6 sec speech as input.
High accuracy, upto 96% under moderate environmental noise.
Like other MLSP turnkey ML solutions, this voiceprint model is shipped in easy deployable format. That is model and API all packaged into single C library, simply link it from your application project and start calling the processing. Currently support following platforms and expanding.
Hifi Fusion F1 DSP
Hifi4 DSP
ARM Cortex M4/7 MCU
For early access please visit www.MLSP.ai product page and get in touch.
Comments