Kay Berkling

Synthesis of Child Speech With HMM Adaptation and Voice Conversion

The synthesis of child speech presents challenges both in the collection of data and in the building of a synthesizer from that data. We chose to build a statistical parametric synthesizer using the hidden Markov model (HMM)-based system HTS, as this technique has previously been shown to perform well for limited amounts of data, and for data collected under imperfect conditions. Six different configurations of the synthesizer were compared, using both speaker-dependent and speaker-adaptive modeling techniques, and using varying amounts of data. For comparison with HMM adaptation, techniques from voice conversion were used to transform existing synthesizers to the characteristics of the target speaker. Speaker-adaptive voices generally outperformed child speaker-dependent voices in the evaluation. HMM adaptation outperformed voice conversion style techniques when using the full target speaker corpus; with fewer adaptation data, however, no significant listener preference for either HMM adaptation or voice conversion methods was found.

Progression in Materials for Learning to Read and Write - a Cross-Language and Cross-Century Comparison of Readers.

This work is part of a larger effort about understanding the effects of didactic materials on acquisition of reading and writing. In this paper, the focus is on progression in primers (beginner readers). Texts are analyzed in terms of complexity, as measured by entropy between letters and phonemes, from the point of view of reading or writing that text. The assumption is that a good teaching text would start low and increase gradually in complexity. At the same time, languages have different requirements on progression depending on their orthographic depth. The goal of this work is to compare readers in various languages for the beginning stages of reading skill acquisition. We show that there is a difference in difficulty across languages and a large span of approaches in primers of approaching this final difficulty.

Introduction to the Special Section on Speaker and Language Recognition

The 16 papers in this special section focus on speaker and language recognition. The advances reported here expand the scope of practical applications for this technology by reducing constraints, such as allowing greater variations in channel conditions.

Automatic Orthographic Error Tagging and Classification for German Texts

Abstract This paper evaluates an automatic spelling error tagger and classifier for German texts. After explaining the existing error tags in detail, the accuracy of the tool is validated against a publicly available database containing around 1700 written texts ranging from first grade to eighth grade. The tool is then applied to a longitudinal study consisting of weekly children’s texts from second and third grades. It can be shown which error categories contribute most significantly to children’s error profiles. Additionally, it can be shown whether or not children make progress on improving in the categories under study.

A Semi-Automatic Word-Level Annotation and Transcription Tool for Spelling Error Categories

In order to train and evaluate tools for the automatic transcription of misspelled texts and automatic annotation of over 20 spelling error categories, it is important to create training data. A very large database of children’s freely written text was collected in the past and in this paper we describe the tool that we have developed in order to manually transcribe and annotate the data. The manual transcription comprises the reconstruction of the orthographically correct word sequence. Annotation is performed on a per-word basis with respect to committed (child spelling) and potential (correct word) spelling error categories. The tool supports human transcribers by suggesting automatically generated annotations. Consistent annotations are propagated and data is presented to the user in a sorted manner to minimize human effort. The tool has been implemented as a web application that makes use of PHP on the server side and a lightweight Java GUI on the client side. The annotated data is stored in a custom made XML schema.

Language identification with inaccurate string matching

We describe a system designed to recognize the language of an utterance spoken by any native speaker over the telephone. The current approach extends our previous work on language identification based on sequences of speech units (K. Berkling and E. Barnard, 1995). To improve performance we extend this work to allow for inaccurate matches of such sequences. Results are reported for distinguishing between English and German. The strength of this algorithm lies in the generalizability from training to test set. We have obtained a means of discriminating between languages based on statistical derivations. Matching sequences inaccurately in a controlled manner allows us to account for variabilities within languages without sacrificing cross language discrimination.