High Performance Embedded Short Time Fourier Transform Architecture for Real-Time Speech Enhancement using Differential Microphone Arrays

Differential Microphone Arrays (DMAs) are devices that exploit the differences in Time of Arrivals (ToA) of the input audio between each sensor in order to separate sources from different directions. Those devices have been used for many different purposes, like audio source localization, speech enhancement, noise suppression etc. DMA is very attractive for processing broadband signals such as speech because of their frequency-invariant beampatterns and small dimension, which make them suitable for the integration into wearable devices such as hearing aids, smartphones and headphones. However, DMA white noise gain (WGN), especially at low frequencies, limits the array performance in terms of obtainable signal-to-noise gain. This limitation can be overcame by processing the audio signal in the Short-Time Fourier transform (STFT) domain at the price of higher complexity of the signal-processing unit. In order to achieve the desired high data throughput for real time applications, the STFT filterbank structure has been implemented in a Cortex M4F embedded CPU using SMID instruction and an embedded Floating Point Unit. The DMA beamforming coefficients can be calculated from the desired beampatterns once offline, stored in system memory and then applied to each STFT frame in order to reduce computational cost. Combining hardware speedups, specialized SMID DSP instructions, high performance FPU and coefficients pre-calculation, the described architecture can achieve 4.096Mbps with only a 8ms of latency, making it suitable for real-time applications.