Abstract:
This work proposes a new approach for speech recognition and voice activity detection tasks. The main limitation of the state-of-the-art solutions for this type of applic...Show MoreMetadata
Abstract:
This work proposes a new approach for speech recognition and voice activity detection tasks. The main limitation of the state-of-the-art solutions for this type of applications is the high power consumption and the silicon area required, which strongly limit their implementation on portable devices. In this manuscript, we propose a new solution based on making use of voltage-controlled-oscillators based analog-to-digital converters (VCO-based ADCs). VCO-based ADCs have enabled efficient audio-oriented implementations in terms of power and area. Here, a VCO-based band-pass filter is proposed to extract the features of analog input sound signals and generate a digital output stream encoding the mel-frequency cepstrum coefficients (MFCCs). This digital signal will be processed afterwards by a decision circuit to determine whether the sound signal corresponds to a pattern. Additionally, the same VCO can be reused to implement both the band-pass filters in the decision mode and the standard ADC output once a keyword or voice has been detected, saving significant area. A behavioral model of the filter was built to validate the performance by simulation. Then, the architecture was designed using a 130-nm CMOS process to get a realistic outlook of the power consumption and the expected occupied area on silicon. Finally, the proposed solution was compared to other equivalent solutions.
Date of Conference: 04-07 August 2019
Date Added to IEEE Xplore: 31 October 2019
ISBN Information: