An Investigation of the Effects of Smoothing and Quantization of the Parameters of Formant-Coded Speech
A. E. Rosenberg, Ronald W. Schafer, L. R. Rabiner · The Journal of the Acoustical Society of America · 1971
Schafer and Rabiner have recently developed an automatic formant and pitch analysis system for connected speech [“System for Automatic Formant Analysis of Voiced Speech,” J. Acoust. Soc. Amer. 47, 634–648 (1970)]. The formant and pitch data obtained are applied to a formant synthesizer at 10-msec intervals with values restricted to ranges appropriate for each parameter. A question of important practical and theoretical interest is the amount of smoothing and quantization of the control data that is permissible without degrading the quality of the resulting synthetic speech. To investigate this question, four all-voiced sentences spoken by male speakers were chosen for an experiment. Stimuli were pro-produced by quantizing or smoothing a specified control parameter over a wide range of values and synthesizing the utterance using the modified control parameters. Subjective evaluation of the synthetic speech was carried out by means of threshold determinations, using a Honeywell DDP-516 laboratory computer. A–B comparison presentations were used in which the reference stimulus was a synthesized version for which the control parameters were unquantized or unsmoothed. Six listeners participated in the experiment. For each experimental parameter, four threshold determinations were obtained from each subject for each of the four sentences. Quantization threshold levels obtained are: pitch period, 3.9 bits; first formant, 2.7 bits; second formant, 3.1 bits. The low-pass-filter cutoff frequency for simultaneous pitch period and formant smoothing was 12 Hz.