FRONICS has developed a proprietary flexible piezoelectric material and applied it to develop ‘Flexible Piezoelectric Acoustic Sensor(f-PAS)’, a VOICE SENSOR with a dramatically improved speaker identification rate. FRONICS’ VOICE SENSOR was made in the form of a piezoelectric thin film in the trapezoidal shape by mimicking the basilar membrane (BM) of the human cochlea. This VOICE SENSOR product can precisely and sensitively acquire voice signals coming through the flexible piezoelectric membrane which resonates for multiple frequency bands of voice.
Unlike conventional microphone sensors, the Flexible Piezoelectric Acoustic Sensor(f-PAS) can be self-powered(always-on) without an external power supply. Each channel can pick up a certain band of voice frequency through a linear response of resonance for corresponding a frequency band, generating voltage when speech is uttered.
While conventional microphone sensors can recognize voice up to about 3 meters away, FRONICS' f-PAS can pick up voices 2 to 4 times that distance.
Further, f-PAS can obtain abundant voice data from one utterance because it receives signals on multiple channels. This significantly reduces errors in speaker identification and voice recognition, painting a picture of the innovative future where the technology will be widely applied to smart home automation, AI assistants, self-driving cars, and biometric authentication security.
In order to verify the performance of FRONICS’ f-PAS in comparison with the conventional microphone sensor, we conducted an experiment to compare the f-PAS with 'SPH****' sensor made by 'Company K', one of the microphone sensors used in 'Company S' mobile phones.
The training databased was created using 70 times of voice utterances by each of the 40 speakers in advance. Then, 7 different voice utterance of the same speakers not included in the training database were used for the actual experiment.
As a result, FRONICS’ VOICE SENSOR reduced the error rate of speaker recognition by 75% compared to the conventional microphone.
Noise Robust Speaker Recognition
– CNN (Deep Learning)
Wolverine – IVA (Conventional)
Algorithm : Independent Vector Analysis (IVA) for source separation
Experimental Condition : 2 speakers + crowd (12 background speakers)
Noise Filtering
– DEMUCS (Deep Learning)
Mixture : voice + alarm noise
Intensity : Noise is 1.8 times higher than voice
VOICE SENSORs are being used in a variety of sectors, creating new profit channels. The core industries that underpin the impending Fourth Industrial Revolution, ranging from big data, IoT, and smart electronics armed with IoT, smart cars, smart devices, and smart home facilities to virtual reality and the defense industry, are tapping into the VOICE SENSOR technology.
KOR