In this paper we propose a memory efficient version of the Gaussian selection (GS) scheme, which is used for speeding up the likelihood calculations of an ASR system. The memory savings are achieved by using non-overlapping (disjoint) clusters instead of the overlapping clusters normally used in GS. As we will show, the new scheme achieves 66% computational savings with a relative increase in word error rate (WER) of 4%. We will also show, that combining the new GS scheme with frame rate reduction and feature masking provides further savings in computation. 75% (4% increase in WER) and 68% (3.5% increase in WER) savings were obtained by adding frame rate reduction and feature masking, respectively

Dictation of natural language text on embedded mobile devices is a challenging task. First, it involves memory and CPU-efficient implementation of robust speech recognition algorithms that are generally resource demanding. Secondly, the acoustic and language models employed in the recognizer require the availability of suitable text and speech language resources, typically for a wide set of languages. Thirdly, a proper design of the UI is also essential. The UI has to provide intuitive and easy means for dictation and error correction, and must be suitable for a mobile usage scenario. In this demonstrator, an embedded speech recognition system for short message (SMS) dictation in US English is presented. The system is running on Nokia Series 60 mobile phones (e.g., N70, E60). The system's vocabulary is 23 thousand words. Its Flash and RAM memory footprints are small, 2 and 2.5 megabytes, respectively. After a short enrollment session, most native speakers can achieve a word accuracy of over 90% when dictating short messages in quiet or moderately noisy environments.

Speech and language processing techniques, such as automatic speech recognition (ASR), text-to-speech (TTS) synthesis, language understanding and translation, will play a key role in tomorrow's user interfaces. Many of these techniques employ models that must be trained using text data. We introduce a novel method for training set selection from text databases. The quality of the training subset is ensured using an objective function that effectively describes the coverage achieved with the strings in the subset. The validity of the subset selection technique is verified in an automatic syllabification task. The results clearly indicate that the proposed systematic selection approach maximizes the quality of the training set, which in turn improves the quality of the trained model. The presented idea can be used in a wide variety of language processing applications that require training with text databases.

Current speech synthesis efforts, both in research and in applications, are dominated by methods based on concatenation of spoken units. New progress in the concatenative text-to-speech (TTS) technology can be made mainly from two directions, either by reducing the memory footprint to integrate the system into embedded system, or by improving the synthesized speech quality in terms of intelligibility and naturalness. In this paper, we are focusing on the memory footprint reduction in a Mandarin TTS system. We show that significant memory reductions can be achieved through duration modeling and memory optimization of the lexicon data. The results obtained in the experiments indicate that the memory requirements of the duration data and lexicon can be significantly reduced while keeping the speech quality unaffected. For practical embedded implementations, this is a significant step towards an efficient TTS engine implementation. The applicability of the approach is verified in the speech synthesis system.

In this paper, two techniques are researched for Jacobian adaptation (JA) in the presence of additive noise. Since the original concept of JA was presented only for static cepstral coefficients, the performance of JA is researched when it is extended to cover also the delta cepstrum. However, this extension or the original concept can not provide accurate recognition performance when the mismatch between the training and recognition environments is out of the linear range of JA. Hence, this problem can be alleviated to some extent by dividing JA into two steps. At first, the adaptation is done e.g. from clean to the target environment having “high” SNR level. After that, the new JA matrixes are calculated and they are used in the second step to adapt the system to the lower target SNR leve1. Both of the above adaptation methods have been compared to cepstral mean normalization (CMN) and parallel model combination (PMC) in isolated word recognition task having a vocabulary of 200 English words. The best performance was achieved with PMC but JA showed comparable performance to CMN and outperformed it when JA was done in two steps from SNR of 25 dB to 5 dB. The system was tested with SpeechDat(II) database by adding noise recorded inside a car to the test set utterances at various SNR levels.

We investigate the technical challenges that are faced when making a transition from the speaker-dependent to speaker-independent speech recognition technology in mobile communication devices. Due to globalization as well as the international nature of the markets and the future applications, speaker independence implies the development and use of language-independent automatic speech recognition (ASR) to avoid logistic difficulties. We propose an architecture for embedded multilingual speech recognition systems. Multilingual acoustic modeling, automatic language identification, and on-line pronunciation modeling are the key features which enable the creation of truly language- and speaker-independent ASR applications with dynamic vocabularies and sparse implementation resources. Our experimental results confirm the viability of the proposed architecture. While the use of multilingual acoustic models degrades the recognition rates only marginally, a recognition accuracy decrease of approximately 4% is observed due to sub-optimal on-line text-to-phoneme mapping and automatic language identification. This performance loss can nevertheless be compensated by applying acoustic model adaptation techniques.

In this paper, we present a multiresolution-based feature extraction technique for speech recognition in adverse conditions. The proposed front-end algorithm uses mel cepstrum-based feature computation of subbands in order not to spread noise distortions over the entire feature space. Conventional full-band features are also augmented to the final feature vector which is fed to the recognition unit. Other novel features of the proposed front-end algorithm include emphasis of long-term spectral information combined with cepstral domain feature vector normalization and the use of the PCA transform, instead of DCT, to provide the final cepstral parameters. The proposed algorithm was experimentally evaluated in a connected digit recognition task under various noise conditions. The results obtained show that the new feature extraction algorithm improves word recognition accuracy by 41 % when compared to the performance of mel cepstrum front-end. A substantial increase in recognition accuracy was observed in all tested noise environments at all different SNRs. The good performance of the multiresolution front-end is not only due to the higher feature vector dimension, but the proposed algorithm clearly outperformed the mel cepstral front-end when the same number of HMM parameters were used in both systems. We also propose methods to reduce the computational complexity of the multiresolution front-end-based speech recognition system. Experimental results indicate the viability of the proposed techniques.

A speaker normalization method using a speech generation model is proposed in order to achieve high-performance speaker adaptation with a small amount of adaptation data. The speaker- and phoneme-dependent vocal tract area function is approximated by the corresponding area function produced by the articulatory model of a standard speaker, combined with phoneme-independent feature quantities of the vocal-tract shape of the normalized target speaker as estimated from the formant frequencies of two vowels. The frequency warping functions are determined from the formant frequencies of speech calculated from the vocal-tract area functions thus obtained, and normalization of the uttered speech is performed by stretching the speech spectrum in the frequency-axis direction. Continuous phoneme recognition experiments using phoneme connection rules show that the recognition error using a gender-dependent model is reduced by about 30% in the proposed method and that recognition performance superior to that of vocal-tract length normalization is obtained. The recognition performance of the proposed method is also equivalent to that of speaker adaptation by moving vector field smoothing (VFS) using 10 phonetically balanced sentences, showing that high-performance speaker adaptation using a small amount of adaptation data can be achieved by the proposed method. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 86(2): 45–56, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.10119

A modified Mellin transform for digital implementation is developed and applied to range radar profiles of naval vessels. The scale invariance property of the Mellin transform provides a means for extracting features from the profiles which are insensitive to the aspect angle of the radar. Past implementations of the Mellin transform based on the FFT have required exponential sampling, interpolation, and the computation of a correction term, all of which introduce errors into the transform. In addition, exponential sampling results in a factor of ln N increase in the number of data points. An alternate implementation, developed in the paper, utilizes a direct expansion of the Mellin integral definition. This direct Mellin transform (DMT) eliminates the implementation problems associated with the FFT approach, and does not increase the number of samples. A scale and translation invariant transform is developed from a modification of the DMT. The MDMT applied to the FFT of the radar profiles results in the desired insensitivity without having the low-pass filtering characteristic that exists in other Fourier-Mellin implementations.