Measurement of signal-to-noise ratio in dysphonic voices by image processing of spectrograms

The measurement of glottal noise was investigated in human and synthesized dysphonic voices by means of two-dimensional (2D) speech processing. A prime objective was the reduction of measurement sensitivities to fundamental frequency (f0) tracking errors and phonatory aperiodicities. An available fingerprint image enhancement algorithm was used for signal-to-noise measurement in narrow band spectrographic images. This spectrographic signal-to-noise ratio estimation method (SNR) creates binary masks, mainly based on the orientation field of the partials, to separate energy in regions with strong harmonics from energy in noisy areas. Synthesized vowels with additive noise were used to calibrate the algorithm, validate the calibration, and systematically evaluate its dependence on f0, shimmer (cycle-to-cycle amplitude perturbation), and jitter (cycle-to-cycle f0 perturbation). In synthesized voices with known signal-tonoise ratios in the 5–40 dB range, SNR estimates were, on average, accurate within ±3.2 dB and robust to variations in f0 (120 Hz or 220 Hz), jitter (0–3%), and shimmer (0–30%). In human /a/ produced by dysphonic speakers, SNR values and perceptual ratings of breathiness revealed a non-linear but monotonic decay of SNR with increased breathiness. Comparison between SNR and related acoustic measurements indicated similar behaviors regarding the relationship with breathiness and immunity to shimmer, but the other methods had marked influence of jitter. Overall, the SNR method did not rely on accurate f0 estimation, was robust to vocal perturbations and largely independent of vowel type, having also potential application in running speech. 2014 Elsevier B.V. All rights reserved.