Guntram Strecha

A multilingual TTS system with less than 1 Mbyte footprint for embedded applications

Text-to-speech (TTS) systems have improved their quality to a large extent lately. This development has resulted in memory requirements of several megabytes that cannot be accepted in many applications, especially in embedded systems. Such applications are usually limited to a footprint of as much as 1 megabyte and require the processing power to be as low as possible. These requirements may be met if the text processing is changed from the usual data-driven algorithms to rule-based processing. Furthermore, the inventory (diphone inventory) should be as small as possible and should be stored in a compressed manner. This is demonstrated by a modified version of the Dresden speech synthesis system, DRESS, which is called microDRESS. Compared to the baseline system, microDRESS does not show essential quality losses apart from the influences of the telephone bandwidth which is appropriate for many embedded applications.

Speech synthesis using HMM based diphone inventory encoding for low-resource devices

In this paper we describe the compression of diphone inventories used by the acoustic synthesis of a concatenative synthesis system. The inventory compression is based on a codebook drawn from the Gaussian mean vectors of phoneme HMMs. There are two encoding/synthesis schemes, a speaker dependent and a speaker independent one. The advantage of the latter is the potential common use of the HM-models by a recognizer and a synthesizer. We describe the steps to encode the inventories as well as the acoustic synthesis using them. Using the proposed method a diphone inventory with 1175 units can be compressed down to 19 kB. We will show that the synthesis quality with HMM-encoded inventories matches the quality of synthesis with AMR- or SPEEX-encoded inventories at noticeably smaller inventory sizes.

Analysis and Synthesis of Glottalization Phenomena in German-Accented English

The present paper investigates the analysis and synthesis of glottalization phenomena in German-accented English. Word-initial glottalization was manually annotated in a subset of a German-accented English speech corpus. For each glottalized segment, time-normalized F0 and log-energy contours were produced and principal component analysis was performed on the contour sets in order to reduce their dimensionality. Centroid contours of the PC clusters were used for contour reconstruction in the resynthesis experiments. The prototype intonation and intensity contours were superimposed over non-glottalized word-initial vowels in order to resynthesize creaky voice. This procedure allows the automatic creation of speech stimuli which could be used in perceptual experiments for basic research on glottalizations.

Analysis-by-synthesis approach for acoustic model adaptation

This paper presents an analysis-by-synthesis approach for acoustic model adaptation. Using artificial speech data for speech recognition systems adaptation, has the potential to address the problem of data sparseness, to avoid speech recordings in real conditions and to provide the capability of performing large number of development cycles for Automatic Speech Recognition (ASR) systems in shorter time. The proposed adaptation framework uses unified ASR and synthesis system to produce artificial adaptation speech signals. In order to confirm the usability of the proposed approach, several experiments were performed where the artificial speech data was coded-decoded by different speech and waveform coders and the acoustic model used for synthesis was adapted for each coder. The recognition results show that the proposed method could be used successfully in the process of speech recognition systems performance assessment and improvement, not only for coded speech effects evaluation and adaptation, but also for other environment conditions.

Cross-Language Acoustic Modeling for Macedonian Speech Technology Applications

This paper presents a cross-language development method for speech recognition and synthesis applications for Macedonian language. Unified system for speech recognition and synthesis trained on German language data was used for acoustic model bootstrapping and adaptation. Both knowledge-based and data-driven approaches for source and target language phoneme mapping were used for initial transcription and labeling of small amount of recorded speech. The recognition experiments on the source language acoustic model with target language dataset showed significant recognition performance degradation. Acceptable performance was achieved after Maximum a posteriori (MAP) model adaptation with limited amount of target language data, allowing suitable use for small to medium vocabulary speech recognition applications. The same unified system was used again to train new separate acoustic model for HMM based synthesis. Qualitative analysis showed, despite the low quality of the available recordings and sub-optimal phoneme mapping, that HMM synthesis produces perceptually good and intelligible synthetic speech.