Search

Homophones pose serious issues for automatic speech recognition (ASR) as they have the same pronunciation but different meanings or spellings. Homophone disambiguation is usually done within a stochastic language model or by an analysis of the homophonous word’s context, similarly to word sense disambiguation. Whereas this method reaches good results in read speech, it fails in conversational, spontaneous speech, where utterances are often short, contain disfluencies and/or are realized syntactically incomplete. Phonetic studies, however, have shown that words that are homophonous in read speech often differ in their phonetic detail in spontaneous speech. Whereas humans use phonetic detail to disambiguate homophones, this linguistic information is usually not explicitly incorporated into ASR systems. In this paper, we show that phonetic detail can be used to automatically disambiguate homophones using the example of German pronouns. Using 3179 homophonous tokens from a corpus of spontaneous German and a set of acoustic features, we trained a random forest model. Our results show that homophones can be disambiguated reasonably well using acoustic features (74% F 1 , 92% accuracy). In particular, this model is able to outperform a model based on lexical context (48% F 1 , 89% accuracy). This paper is of relevance for speech technologists and linguists: amodule using phonetic detail similar to the presented model is suitable to be integrated in ASR systems in order to improve recognition. An approach similar to the work here that combines the automatic extraction of acoustic features with statistical analysis is suitable to be integrated in phonetic analysis aiming at finding out more about the contribution and interplay of acoustic features for functional categories. [ABSTRACT FROM AUTHOR]

This paper presents GRASS (Graz corpus of Read and Spontaneous Speech), the first large scale speech database for Austrian German with both read and conversational speech. In total, the corpus contains approximately 1900 min of speech in which 38 speakers produced more than 220,000 word tokens from 14,593 different word types. The corpus consists of three components. First, the Conversational Speech Component contains free conversations of one hour length between friends, colleagues, couples, or family members. Second, the Commands Component contains commands and keywords which were either read or elicited by pictures. Third, the Read Speech Component contains phonetically balanced sentences and digits. The speech of all components has been recorded at fullband quality in a soundproof recording-studio with head-mounted microphones, large-diaphragm microphones, a laryngograph, and with a video camera. The corpus was fully annotated at the orthographic level, and partly also at the segmental, sub-segmental and prosodic level. Our analysis of conversational speech characteristics such as overlapping speech, laughter, repetitions, hesitations and the use of colloquial and dialectal words allows us to conclude that the conversational speech material is highly casual in nature. The collected corpus provides conversational material for phoneticians and linguists interested in topics specific for Austrian German (e.g., pronunciation variability, prosody, syntax of spoken Austrian German), and for those studying talk in interaction in general (turn-taking, grounding, entrainment, extra-linguistic factors etc.). Furthermore, it is a valuable resource for speech technologists interested in the development of ASR and dialogue systems for different speaking styles of Austrian German. [ABSTRACT FROM AUTHOR]

With the growing interest among speech scientists in working with natural conversations also the popularity for using articulatory-acoustic features as basic unit increased. They showed to be more suitable than purely phone-based approaches. Even though the motivation for AF classification is driven by the properties of conversational speech, most of the new methods continue to be developed on read speech corpora (e.g., TIMIT). In this paper, we show in two studies that the improvements obtained on read speech do not always transfer to conversational speech. The first study compares four different variants of acoustic parameters for AF classification of both read and conversational speech using support vector machines. Our experiments show that the proposed set of acoustic parameters substantially improves AF classification for read speech, but only marginally for conversational speech. The second study investigates whether labeling inaccuracies can be compensated for by a data selection approach. Again, although an substantial improvement was found with the data selection approach for read speech, this was not the case for conversational speech. Overall, these results suggest that we cannot continue to develop methods for one speech style and expect that improvements transfer to other styles. Instead, the nature of the application data (here: read vs. conversational) should be taken into account already when defining the basic assumptions of a method (here: segmentation in phones), and not only when applying the method to the application data [ABSTRACT FROM AUTHOR]

Abstract: In spontaneous, conversational speech, words are often reduced compared to their citation forms, such that a word like yesterday may sound like . The present chapter investigates such acoustic reduction. The study of reduction needs large corpora that are transcribed phonetically. The first part of this chapter describes an automatic transcription procedure used to obtain such a large phonetically transcribed corpus of Dutch spontaneous dialogues, which is subsequently used for the investigation of acoustic reduction. First, the orthographic transcriptions were adapted for automatic processing. Next, the phonetic transcription of the corpus was created by means of a forced alignment using a lexicon with multiple pronunciation variants per word. These variants were generated by applying phonological and reduction rules to the canonical phonetic transcriptions of the words. The second part of this chapter reports the results of a quantitative analysis of reduction in the corpus on the basis of the generated transcriptions and gives an inventory of segmental reductions in standard Dutch. Overall, we found that reduction is more pervasive in spontaneous Dutch than previously documented. [Copyright &y& Elsevier]