The effect of language identification accuracy on speech recognition accuracy of proper names

Abstract

Utilising the known language of origin of a name can be useful when predicting the pronunciation of the name. When this language is not known, automatic language identification (LID) can be used to influence which language-specific grapheme-to-phoneme (G2P) predictor is triggered to produce a pronunciation for the name. We investigate the implications when both the LID system and the G2P system generate errors: what influence does this have on a resulting speech recognition system? We experiment with different approaches to LID-based dictionary creation and report on results in four South African languages: Afrikaans, English, Sesotho and isiZulu.

Bilateral G2P accuracy: Measuring the effect of variants

“…In earlier work [2], we observed that G2P accuracy and ASR performance do not always correlate well if pronunciation variants occur. Also, we noticed that the way pronunciation variants are dealt with when measuring accuracy has a significant effect on the result.…”

“…Ambiguity arises when standard G2P accuracy measures are used to score variants; that is, measuring predicted pronunciation variants against the gold standard reference dictionary that also includes variants. In some of our earlier work, little correlation was found between G2P accuracy and ASR performance in such cases [2]. One apparent reason for artificially improved G2P accuracy, depending on measure selected, is that either over-generation or under-generation of variants is unfairly advantaged.…”

“…Estimating the G2P accuracy of variant-based dictionaries requires an approach that allows penalising over-or undergeneration of pronunciation variants. In order to obtain an empirical result for the task of balancing between G2P accuracy and ASR performance, our experimental setup is designed to mimic previous work carried out in [2]. We used the same setup as discussed in [2].…”

“…In their work, a data-driven approach was employed to observe pronunciation variation at syllable level. For each word, a syllable-level pronunciation was obtained using the CMU pronunciation lexicon 1 and NIST syllabification software 2 . Their experiment was performed on the TIMIT corpus 3 of read speech.…”

“…It provides a mapping from orthographic to phonetic based on the ARPAbet phoneme set [5]. 2 Software that marks syllable boundaries and its implementation is based on Kahn's theory of syllabification [6]. 3 An acoustic-phonetic continuous speech designed for the development and evaluation of automatic speech recognition systems based on their posterior probability.…”