Measuring Speech Quality of Laptop Microphone System using PESQ

Abstract: Signal compression algorithms are required in speech processing system to support the limited bandwidth requirements. And while compressing the signal, speech quality must not be degraded and modern speech methods makes use of perceptual information present in the signal to distinguish which part of the signal needs to be sent and which needs to be discarded. Measures like Signal to Noise Ratio (SNR) do not provide accurate measurement of speech quality of these modern communication systems. Traditionally, Mean opinion score(MOS) is an subjective quality measure with panel of listeners and conducting these tests is very expensive and time consuming making it impractical for all test conditions. An objective quality measure Perceptual evaluation of speech quality was recommended by ITU-T as an enhanced speech quality measure, especially in a laptop microphone system. This research includes the elimination of fan noise, keystroke noise and measurement of speech quality for different noises like factory noise, engine noise etc using PESQ at different SNRs varies from 0db to 45db. The speech enhancement algorithms like Wiener Beamformer, spectral subtraction and combination of Wiener Beamformer and Spectral Subtraction is implemented in Matlab environment, PESQ measures the speech quality range from 1 to 4.5, 1 means poor quality and 4.5 means excellent. In this thesis simulation results shows that the PESQ values for different noise (fan noise, keystroke, factory noise and engine noise) signals using different speech enhancement algorithms (systems) for various SNRs like 0db to 45db. Using these speech enhancement algorithms reduce the fan noise, keystroke noise, factory noise and engine noise for various SNRs with an SNRI and maintaining speech quality. In the simulation results of evaluation method-1, in which the pure speech signal is compared with itself, then PESQ score is observed as 4.5 on the scale of 4.5. In the evaluation method-2, the different noise signals with different sampling frequencies i.e fs=8K and 16K are tested with pure speech signal, to analyze the PESQ variation over different sampling frequencies. With the obtained results it shows that for fs=8kHz, the PESQ score is higher when compared to fs=16KHz as the SNR is gradually increasing from 0 to 45dB. In the evaluation method-3, the Wiener Beamformer system itself preserves the speech quality while attenuating different noises i.e. PESQ score is higher, when compared to SS system and combined system.