Stefan Rapp

Generating non-native pronunciation variants for lexicon adaptation

Abstract Handling non-native speech in automatic speech recognition (ASR) systems is an area of increasing interest. The majority of systems are tailored to native speech only and as a consequence performance for non-native speakers often is not satisfactory. One way to approach the problem is to adapt the acoustic models to the new speaker. Another important means to improve performance for non-native speakers is to consider non-native pronunciations in the dictionary. The difficulty here lies in the generation of the non-native variants, especially if various accents are to be considered. Traditional approaches to model pronunciation variation either require phonetic expertise or extensive speech databases. They are too costly, especially if a flexible modelling of several accents is desired. We propose to exclusively use native speech databases to derive non-native pronunciation variants. We use an English phoneme recogniser to generate English pronunciations for German words and use these to train decision trees that are able to predict the respective English-accented variant from the German canonical transcription. Furthermore we combine this approach with online, incremental weighted MLLR speaker adaptation. Using the enhanced dictionary and the speaker adaptation alone improved the word error rate of the baseline system by 5.2% and 16.8%, respectively. When both methods were combined, we achieved an improvement of 18.2%.

The Dynamic Lexicon

The dynamic lexicon is one of the central knowledge sources in SmarTkom that provides the whole system with the capabability to dynamically update the vocabulary. The corresponding multilingual pronunciations, which are needed by all speech-related components, are automatically generated.

Using Rough Sets Theory to Predict German Word Stress

The location of primary stress in German morphologically simple words is an interesting domain for the application of symbolic machine learning ML techniques It is commonly accepted among phonologists that stress location is predictable from the syllable structure and segmental setup of the last syllables in the word Using a rough sets based classi er we test some assumptions about the phonological factors underlying stress placement Whereas phonologists have arrived at these assumptions mostly in a deductive way machine learning pro vides the opportunity for corpus based empirical evaluation The words included in our corpus are nouns that are monomorphemic or derived by a nonnative su x and they contain more than one non schwa vowel They are drawn from a textbook on German word stress and can thus be seen as representative to the stress assignment task All the words have been transcribed phonetically according to the DUDEN pronouncing dictionary along with syl lable boundaries added by the authors With pronunciation variants the corpus consists of cases di erent words Two types of coding have been com putationally extracted from the corpus one being the information commonly used by phonologists the other a more detailed set of attributes consisting of presence of syllable onset vowel height vowel length vowel tenseness and num ber of consonants in syllable coda The decision attribute is the location of stress on either the nal the penultimate or the antepenultimate syllable In three experiments we used the rough sets based commercially available ML Tool DatalLogic R to induce symbolic rules from the attributes Our rst experiment shows that it is valid to concentrate on the last three syllables of the word Using a learn test scenario we found that for both types of coding the inclusion of information beyond three syllables from the right only makes a negligible contribution to the accuracy of the induced rules Within a learn test setting the second experiment demonstrates that the more detailed type of coding yields substantially improved predictive accuracy over the phonological coding In a third experiment we found that this improvement is due to only the vowel length attributes In general our experiments have demonstrated that ML techniques are a useful tool in the investigation of current issues in linguistic research