Evaluation of an automated formant estimation procedure with optimized formant ceiling

Sammanfattning: This study evaluates an automated formant estimation procedure designed to adapt to speakers and variations in speech. The adaption is achieved by using the formant ceiling with the least variation (in combined estimates of F1 and F2) as the optimal ceiling. This optimization renders the best possible estimations given the data, therefore it could presumably also adapt to variations such as high fo. The procedure has not been evaluated by using material with known formant frequencies. Therefore, this is done here. The performance of the procedure is tested through comparison with a common procedure with fixed ceilings, based on speaker sex. The estimations are carried out on synthetic vowel tokens, systematically varied in formant frequencies and in fo, to match the natural variation within vowels and between speakers. The formant estimations are compared to target values, compared between procedures and to earlier studies. The results reveal that the formant estimation procedure with optimized ceilings does not perform better than the common procedure. Both procedures perform better than earlier methods, but neither deals satisfactorily with high fo.