Formant measurement errors: Preliminary results from synthetic speech
Philip Harrison, J P French Associates · 2007
This presentation introduces the preliminary results from the latest phase of work on the variability of formant measurements. The initial study (Harrison, 2004) and subsequent IAFPA conference presentations (Harrison, 2005 & 2006) examined the magnitude of the variation in formant measurements caused by altering analysis settings. Formant measurements were made using the automatic formant tracker in three different programs whilst analysis parameters were systematically varied. The source material was a word list read by two male speakers recorded simultaneously via a microphone and at the distant end of a telephone line. The main limitation of the study was that owing to the nature of formants it was not possible to know what their true values were. The consequence being that it was only possible to consider the variation in relative terms rather than as absolute errors. The current phase of work overcomes this problem by using synthesised speech as the source material. The speech is synthesised using the source-filter method. This allows the actual formant values to be specified in the synthesis process. A further limitation of the initial study was that the word list contained multiple instances of only 5 different vowels. Even though they represented the extremes of the vowel space and a central vowel they did not provide comprehensive coverage of the whole space. The use of a speech synthesiser