Fig. 7

Evaluation procedure, in which a comparison on the output of all forced aligners to a hand-aligned ground-truth, is performed. However, metrics should be calculated considering a one-to-one mapping between inference and reference phonemes. This is where the M2M mapping is applied to make different phone sets inside TextGrid files match the SAMPA phonetic alphabet version used by FalaBrasil’s G2P converter, which is provided by the lexicon generated over transcriptions from the evaluation speech corpora for both male and female speakers (lex M/F). That way, M2M mapping ensures each reference and inference utterance shares the same number of phones before undergoing evaluation