Automatic phonetic baseform determination

Abstract

Phonetic baseforms are the basic recognition units in most large vocabulary speech recognition systems. These baseforms are usually determined by hand once a vocabulary is chosen and not modified thereafter. However, many applications of speech recognition, such as dictation transcription, are hampered by a fixed vocabulary and require the user be able to add new words to the vocabulary. At least one phonetic baseform must be assigned to each new word to properly integrate the word into the recognition system. Dictionary lookup is often unsuccessful in determining a phonetic baseform because new words are often names or task-specific jargon; also, talkers tend to have idiosyncratic pronunciations for a substantial fraction of words. This paper describes a series of experiments in which the phonetic baseform is deduced automatically for new words by utilizing actual utterances of the new word in conjunction with a set of automatically derived spelling-tosound rules. We evaluated recognition performance on new words spoken by two different talkers when the phonetic baseforms were extracted via the above approach. The error rates on these new words were found to be comparable to or better than when the phonetic baseforms were derived by hand, thus validating the basic approach.

Learning Lexicons From Speech Using a Pronunciation Mixture Model

“…Of particular interest to this work, are instances in which spoken examples are used to refine pronunciations [25]- [29]. The work of [27], for example, deduces a pronunciation given a word or grapheme sequence and an utterance of the spoken word . This research uses a decision tree to model which was later shown to produce poor results when compared to graphone models on L2S tasks.…”

Narrow Adaptive Regularization of weights for grapheme-to-phoneme conversion

“…Out-of-vocabulary (OOV) words are the bottleneck in largevocabulary open-domain speech recognition systems [1] and text-to-speech systems [2]. In order to solve the problem of OOV words, Grapheme-to-phoneme (g2p) conversion, which is structured learning problems for which there are an extremely large number of candidate answers, has been used for a long time.…”

A turbo-style algorithm for lexical baseforms estimation

“…There has been significant research on automatic lexical generation [5,6,7]. However, the novel contribution of this work is two-fold: (1) Spoken examples of both the spelling and the word are used as opposed to the word only, and (2) a bi-directional L2S model is used to exchange bias information between the spelling and pronunciation domain so as to boost the overall performance of the tandem model.…”