Discriminative pronunciation learning for speech recognition for resource scarce languages

Abstract

In this paper, we describe a method to create speech recognition capability for small vocabularies in resource-scarce languages. By resource-scarce languages, we mean languages that have a small or economically disadvantaged user base which are typically ignored by the commercial world. We use a high-quality welltrained speech recognizer as our baseline to remove the dependence on large audio data for an accurate acoustic model. Using cross-language phoneme mapping, the baseline recognizer effectively recognizes words in our target language. We automate the generation of pronunciations and generate a set of initial pronunciations for each word in the vocabulary. Next, we remove potential conflicts in word recognition by discriminative training.

Cross-language Phoneme Mapping for Low-resource Languages: An Exploration of Benefits and Trade-offs

“…Our work builds directly on previous work, namely the Speechbased Automated Learning of Accent and Articulation Mapping (SALAAM) [12,11] algorithm, as implemented in the open source tool Lex4All [14,13].…”

“…The primary idea behind the SALAAM technique is to find the best pronunciation sequence for a given word in a target language from one or more audio samples by using a source language speech recognizer to perform phone decoding (decoding by phoneme) [10]. Since most commercial speech recognizers do not directly support phone decoding, the SALAAM technique uses a specially-designed grammar to mimic phonedecoding [10,13]. This is achieved by creating a recognition grammar representing a phoneme super wildcard to guide pronunciation discovery.…”

“…Previous studies have only used English (US) and French (France) as the source languages, with a maximum of 10 alternative pronunciation per target language word type [10,12,11,13]. We investigate the performance and tradeoffs of different source languages and training techniques using four source languages: English (US), French (France), German (Germany) and Mandarin (China), and five target languages: English (South African), Sotho, Afrikaans, chiShona and Kiswahili.…”

“…These pronunciation maps could be handwritten, but they demand the use of an expert linguist who is fluent in both the source and target languages but often do not yield high recognition accuracy [10]. Therefore, processes of automatically creating cross-language phoneme mappings between languages were developed [12,13,14], eliminating the requirement of an expert linguist. This technique has been used in a number of information, communication, and technology for development (ICT4D) projects in the health sector [4], agriculture sector [8,6] and for research purposes [14,10].…”

Artificial Neural Network-Based Speech Recognition Using Dwt Analysis Applied On Isolated Words From Oriental Languages

“…If the frequency range of an input speech signal is from 0 to f, then the frequencies above (f÷2)are cut off by the low-pass filter, while the frequencies below (f÷2)are cut off by the high-pass filter. Filtering a speech signal is equivalent to the convolution mathematical operation of the signal [22][23] by the filter's impulse response (see Eq. (10) and Eq.…”