Analysis: Processor mimics ear, gets 25 dB SNR gain

The recent Hot Chips Conference held at Stanford University had a number of papers of interest to the digital signal processing community. Of particular interest was a paper from Audience Inc. titled "Voice Processor Based on the Human Hearing System." This paper explains how the company's A1010 processor improves audio SNR by 25 dB (typical) in applications like cell phones.

The key innovation in the A1010 is its ability to handle non-stationary noise. The A1010 achieves this feat through signal processing based on the human auditory system. One key piece of this signal processing is the Fast Cochlea Transform (FCT). The FCT has a number of properties which distinguish it from the conventional FFT (Figure 1).

Figure 1. The Fast Cochlea Transform (FCT).

Engineers involved in video and image compression may see the similarities between the FCT behavior and the properties of Wavelet functions when used as basis functions for compression. The logarithmic response is also reminiscent of the Cepstrum techniques used in Homomorphic DSP.