Vocal Tract Normalization Based on Formant Positions

This paper presents our initial results in a new approach to vocal tract normalization (VTN). In experiments based on continuous automatic speech recognition (ASR) the VTN procedure is in general carried out in both training and test phase. In the training phase it is used to obtain speaker independent acoustic models of phones. In the test phase it is used to convert input observations into observations nearer to the ones corresponding to the universal speaker. The approach described in this paper is new, because instead of training a single set of acoustic models for the universal speaker, several sets of acoustic phone models corresponding to speakers with similar vocal tract lengths were created. Instead of using the VTN procedure in the test phase, the recognized sequence estimated as the most likely one among sequences based on different acoustic model sets was identified as the final recognition result. Normiranje vokalnega trakta na podlagi lege formantov V prispevku so predstavljeni začetni rezultati novega pristopa k normiranju vokalnega trakta (NVT). V eksperimentih, ki temeljijo na samodejnem razpoznavanju tekočega govora, se postopek NVT izvaja tako v učni kot v testni fazi. V učni fazi se uporablja za pridobivanje akustičnih modelov fonov, ki niso odvisni od govorca. V testni fazi se uporablja za pretvorbo vhodnih opažanj v opažanja, ki so bližja tistim, ki ustrezajo univerzalnemu govorcu. Pristop, ki je opisan v tem prispevku, je nov: namesto da bi učili posamezno množico akustičnih modelov za univerzalnega govorca, je bilo ustvarjenih več množic akustičnih modelov fonov, ki ustrezajo govorcem s podobno dolžino vokalnega trakta. Namesto uporabe postopka NVT v testni fazi je bil končni rezultat razpoznavanja prepoznano zaporedje, ki je bilo ocenjeno kot najbolj verjetno med zaporedji, temelječimi na različnih množicah