Four open source formant trackers, three LPC-based and one based on Deep Learning, were evaluated on the same American English data set VTR-TIMIT. Test data were time-synchronized to avoid differences due to different unvoiced/voiced detection strategies. Default output values of trackers (e.g. producing 500Hz for the first formant, 1500Hz for the second etc.) were filtered from the evaluation data to avoid biased results. Evaluations were performed on the total recording and on three American English vowels [i:], [u] and [ʌ] separately. The obtained quality measures showed that all three LPC-based trackers had comparable RSME error results that are about 2 times the inter-labeller error of human labellers. Tracker results were biased considerably (in average too high or low), when the parameter settings of the tracker were not adjusted to the speaker's sex. Deep Learning appeared to outperform LPC-based trackers in general, but not in vowels. Deep Learning has the disadvantage that it requires annotated training material from the same speech domain as the target speech, and a trained Deep Learning tracker is therefore not applicable to other languages.
@InProceedings{SCHIEL18.28, author = {Florian Schiel and Thomas Zitzelsberger}, title = "{Evaluation of Automatic Formant Trackers}", booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)}, year = {2018}, month = {May 7-12, 2018}, address = {Miyazaki, Japan}, editor = {Nicoletta Calzolari (Conference chair) and Khalid Choukri and Christopher Cieri and Thierry Declerck and Sara Goggi and Koiti Hasida and Hitoshi Isahara and Bente Maegaard and Joseph Mariani and Hélène Mazo and Asuncion Moreno and Jan Odijk and Stelios Piperidis and Takenobu Tokunaga}, publisher = {European Language Resources Association (ELRA)}, isbn = {979-10-95546-00-9}, language = {english} }
Summary of the paper