G2p variant prediction techniques for ASR and STD

Abstract

Introducing pronunciation variants into a lexicon is a balancing act: incorporating necessary variants can improve automatic speech recognition (ASR) and spoken term detection (STD) performance by capturing some of the variability that occurs naturally; introducing superfluous variants can lead to increased confusability and a decrease in performance. We experiment with two very different grapheme-to-phoneme variant prediction techniques and analyze the variants generated, as well as their effect when used within fairly standard ASR and STD systems with unweighted lexicons. Specifically, we compare the variants generated by joint sequence models, which use probabilistic information to generate as many or as few variants as required, with a more discrete approach: the use of pseudophonemes within the default-and-refine algorithm. We evaluate results using three of the 2013 Babel evaluation languages with quite different variant characteristics -Tagalog, Pashto and Turkish -and find that there are clear trends in how the number and type of variants influence performance, and that the implications for lexicon creation for ASR and STD are different.

The South African directory enquiries (SADE) name corpus

“…Also, we employed log-probability voting to select the final language of origin of a word as described in [14]. Training data was obtained from the NCHLT-inlang [13,5] dictionaries, providing 15,000 unique words per language. Data was preprocessed by removing words with fewer than two characters, converting all characters to lowercase and removing any loan words that are easily identifiable based on any foreign characters included in the word.…”

Bilateral G2P accuracy: Measuring the effect of variants

“…The G2P accuracy for variant-based dictionaries is analysed using four metrics defined in [1], namely variantbased phone accuracy (V-PA), variant-based word accuracy (V-WA), single-best phone accuracy (S-PA) and single-best word accuracy (S-WA). (Also see [2].…”

“…The dictionary is typically developed manually, using human expertise, or created in a data-driven manner from a training sample. Capturing acoustic variabilities, such as dialect, differences in semantics or accents, could result in incorporating different pronunciation variants into the dictionary [1]. Capturing pronunciation variation in a dictionary is controversial as, in practice, it can either increase or decrease ASR performance.…”

“…Jouvet et al [8] evaluated the average number of pronunciation variants per word that is required to optimise recognition accuracy in the context of French broadcast news data. Given a range of thresholds, pronunciation variants were selected 1 An open source pronunciation dictionary created for speech recognition and synthesis research. It provides a mapping from orthographic to phonetic based on the ARPAbet phoneme set [5].…”

“…For variant-based dictionaries, that is, when measuring predicted pronunciation variants against the gold standard reference dictionary that also includes variants, it is not fully straightforward how variants should be accounted for. Each reference variant can be treated as a separate word, or as in [1], the variants linked to their producing word. The two alternative measures used here were referred to as single-best and variant-based versions of WER and PER.…”