To examine the robustness of Acoustic-Feature definitions, broad-phonetic label alignment was carried out on Danish, English and Italian continuous speech recordings after training a Neural-Net-based system on speech material taken from more than one language. Shared feature specifications for phonemes adjudged to be sufficiently similar across languages increased the training base for those sounds. Other phonemes with no equivalents across the languages were trained only on language specific material. In some cases labelling accuracy was better than earlier experiments in which only language specific training was carried out. However, results indicate that the amount of training material alone does not increase accuracy. Contextual variation in the segment transitions makes some sounds inherently more difficult to label automatically.
Bibliographic reference. Dalsgaard, Paul / Andersen, Ove / Barry, William (1991): "Multi-lingual acoustic-phonetic features for a number of european languages", In EUROSPEECH-1991, 685-688.