23 December 2022 in IEEE Journal of Solid-State Circuits
This work introduces a time-interleaved architecture to tackle the speed-resolution bottleneck of the noise-shaping (NS) successive approximation register (SAR) quantizer in a continuous-time (CT) sigma-delta modulator (SDM). A critical insight is that introducing a delay in the NS quantizer feedback loop enables complete parallelization of the NS quantizer operations. The extra time from parallelization greatly relaxes loop filtering and residue integration and enables a high quantizer resolution. Furthermore, the extra time in the feedback loop allows data weighted averaging (DWA), which eliminates the need for calibration. The prototype comprises a second-order CT front-end and a fully interleaved first-order NS quantizer. The prototype is fabricated in a 28 nm process, occupies 0.072 mm2 and consumes 6.4 mW. The peak signal-to-noise ratio (SNR), signal-to-noise-distortion ratio (SNDR), and dynamic range (DR) are 83.9, 81.6, and 84.8 dB, respectively, in a 15.625 MHz bandwidth. The corresponding Schreier SNDR figure of merit (FoM) is 175.5 dB.
24 August 2021 in IEEE Journal of Solid-State Circuits
Beamforming is an essential tool for speaker selection and rejection of environmental noise in automatic speech recognition. This work harnesses the efficiency of delay-and-sum (DAS) beamforming by combining it with constant-directivity beamforming (CDB) and frequency-domain feature extraction. CDB facilitates DAS by restricting the bandwidth for different microphone configurations. An array of sigma–delta modulators (SDMs) digitizes eight microphone inputs. The design takes advantage of bitstream processing of the modulator outputs for beamforming and extracting 60 Mel spectrum power features. The prototype device is fabricated in the 40-nm CMOS and occupies 1.1 mm2. Each SDM consumes 91 mW and has a measured signal-to-noise and distortion ratio of 84 dB for an 8-kHz bandwidth. The beamformer and feature extractor consume a dynamic power of 76 and 122 mW, respectively. The entire power consumption of the prototype is 3.95 mW, including leakage power. Processing the Mel spectrum outputs with a DNN, the keyword spotting accuracy in the presence of noise improves from 74% without beamforming to 93% with beamforming.
9 December 2024 in IEEE Open Journal of the Solid-State Circuits Society
This article introduces a fully dynamic SAR-assisted pipeline ADC that uses a floating ring amplifier (FLORA) and gain-enhancing Miller negative capacitance (Miller negative-C). FLORA is a fully dynamic and bias-free ring amplifier powered by reservoir capacitors. Different reservoir capacitors for auto-zero and amplification phases optimize the power consumption and dominant pole locations. Furthermore, FLORA enhances speed without needing common-mode load capacitors in each stage and does not need a switched-capacitor common-mode feedback circuit at the output. The Miller negative-C improves the accuracy of the closed-loop residue amplifier by reducing the gain error related to the product of the finite operational amplifier gain and the feedback factor. This gain error compensation scheme eliminates the need for extra circuitry or correction phases. It occupies a small area and requires little additional power consumption. The paper analyzes the stability, effective range, and settling behavior of an amplifier with Miller negative-C. The prototype ADC implemented in a 28-nm CMOS process achieves an SNDR and an SFDR of 67.9 dB and 84.3 dB, respectively, while consuming 1.72 mW at 150 MS/s from a 1 V supply. The corresponding Walden and Schreier SNDR FoMs are 5.7 fJ/conversion-step and 173 dB, respectively.
23 November 2023 in IEEE Journal of Solid-State Circuits
This work introduces a speech recognition frontend system that solves the problems of conventional adaptive beamforming (ABF) with: 1) low digital signal processing (DSP) power consumption (3 $\times $ lower than state-of-the-art ABF) thanks to an innovative greedy blocking matrix (GBM) employing simple calculations; 2) automatic direction-of-arrival (DOA) error compensation with direction tracking delay-and-sum beamformer aided by the GBM; 3) a multimode hybrid analog-to-digital converter (ADC) adapts to signal conditions; and 4) multimode beamforming takes advantage of high-signal signal-to-noise ratio (SNR) to reduce total power by 54%. A prototype fabricated in a 40 nm process occupies 0.94 mm2 while consuming 157 and 72 $\mu \text{W}$ in high-power and low-power modes, respectively. The proposed system improves speech recognition accuracy from 54% to 83% under noisy conditions.
1 September 2021 in IEEE Solid-State Circuits Letters
A speech recognition front-end combines a four-channel adaptive beamformer and a 40-feature Mel frequency extractor. The prototype processes the bitstream outputs of third-order delta-sigma modulators with a robust generalized sidelobe canceller (RGSC) for accurate steering. For a given steering vector, the beamformer adaptively places a null in the direction of the noise. Hardware sharing and DSP clock optimization reduce area and power consumption. The prototype is fabricated in 40-nm CMOS and occupies an active area of 0.89 mm 2 . The prototype beamformer improves speech recognition accuracy in noisy conditions from 64% to 90%.