126 An Approach to Optimize the Coding of a Formant Synthesizer Parameters FREQUENZ 39 (1985) 5 An Approach to Optimize the Coding of a Formant Synthesizer Parameters Angenäherte Optimalcodierung der Parameter bei der Formantsprachsynthese By Selim S. Awad* and Bernard Guerin** Abstract: An optimum coding of the parameters of a formant speech synthesizer has been found. The optimization of these control parameters is based on statistical and subjective criteria. The synthesizer used is a parallel synthesizer comprising 19 independant parameters. The utterances chosen for experimentation are groups of high quality synthetic French CVCV utterances. The proposed procedure consists of three steps. The first is a statistical study carried out on the control parameters of the synthetic utterances in order to find the optimum dynamic range of each parameter. In the second step, the minimum number of bits necessary for quantizing each parameter is found. The third and final step is to optimize the sampling rate of each interval in each utterance. Übersicht: Es wird eine angenäherte Optimalcodierung der Parameter bei der Formantsprachsynthese angegeben. Die Optimierung dieser Kontrollparameter beruht auf statistischen und subjektiven Kriterien. Der benutzte Synthesierer ist ein Parallelsynthesierer mit 19 unabhängigen Parametern. Die für die Untersu- chungen gewählten Wörter sind Gruppen synthetischer CVCV-Sprachbausteine hoher Qualität in Französisch. Das vorgeschlagene Verfahren arbeitet in drei Stufen..Die erste ist eine statistische Untersuchung an den Kontrollparametern der synthetischen Bausteine, um den optimalen Dynamikbe- reich jedes Parameters zu finden. Im zweiten Schritt wird die kleinste Anzahl von Bits gefunden, die zur Quantisierung jedes Parameters notwendig ist. Der dritte und letzte Schritt dient zur Optimierung der Abtastfrequenz für jedes Intervall in den Sprachbausteinen. Für die Dokumentation: Formantsprachsynthese / Optimalcodierung / Statistisches Verfahren / Optimalquantisierung 1. Introcuction Formant synthesizers are considered to be able to pro- duce high quality synthetic speech. This is due to the fact that the controlling parameters of the synthesizer can be independantly controlled in order to match the spectral characteristics of the synthetic speech with the natural speech. A disadvantage of the formant speech synthesis is the relatively high bit-rate necessary for coding its para- meters. In this paper we propose an optimization pro- cedure aiming at reducing the total bit-rate. The first step in our procedure is to determine the effective dynamic ranges of the parameters, since a larger than necessary dynamic range of a parameter increases the quantization error for the same number of bits necessary for quantizing this parameter. For the French language, we can find statistical studies on the effective dynamic ranges of the formant frequencies and the fundamental frequency [1]. Concerning the other parameters, precisely the gain para- meters and the bandwith of the formants, the statistical study must be carried out on a relatively large number of utterances synthesized by the chosen synthesizer. In the second step of our procedure, we determine the minimum number of bits necessary to quantize each para- meter with no noticeable degradation in the quality of the synthetic speech. This step is composed of two sub-steps. The first sub-step is to find the minimum number of bits necessary to quantize each parameter independantly, while the second step is to find the minimum number of bits when all the parameters are quantized simultaneously. CHOICE OF THE CORPS UTTERANCES (1) VOICED FRICATIVES (2) VOICED STOPS DETERMINATION OF THE USEFUL DYNAMIC RANGES OF THE PARAMETERS DETERMINATION OF THE THRESHOLD NUMBER OF BITS AND THE TYPE OF QUANTIZATION (LINEAR OR LOGARITHMIC) (1) INDEPENDANT QUANTIZATION (2) SIMULTANEOUS QUANTIZATION * University of Michigan-Dearborn, Dept. of Electrical Engineering, USA. ** Institut de la Communication Parlee, E.N.S.E.R.G., Grenoble (France). OPTIMIZATION OF THE SAMPLING INTERVAL (VARIABLE SAMPLING INTERVALS) Fig. 1: The proposed optimization procedure Brought to you by | New York University Bobst Library Technical Services Authenticated Download Date | 6/1/15 5:28 PM