This paper presents methods for speaker adaptive modeling using vocal tract normalization (VTN) along with experimental tests on three databases. We propose a new training method for VTN: By using single-density acoustic models per HMM state for selecting the scale factor of the frequency axis, we avoid the problem that a mixture-density tends to learn the scale factors of the training speakers and thus cannot be used for selecting the scale factor. We show that using single Gaussian densities for selecting the scale factor in training results in lower error rates than using mixture densities. For the recognition phase, we propose an improvement of the well-known two-pass strategy: by using a non-normalized acoustic model for the first recognition pass instead of a normalized model, lower error rates are obtained. In recognition tests, this method is compared with a fast variant of VTN. The two-pass strategy is an efficient method, but it is suboptimal because the scale factor and the word sequence are determined sequentially. We found that for telephone digit string recognition this suboptimality reduces the VTN gain in recognition performance by 30% relative. In summary, on the German spontaneous speech task Verbmobil, the WSJ task and the German telephone digit string corpus SieTill, the proposed methods for VTN reduce the error rates significantly.

This paper presents improved methods for vocal tract normalization (VTN) along with experimental tests on three databases. We propose a new method for VTN in training: by using acoustic models with single Gaussian densities per state for selecting the normalization scales the need for the models to learn the normalization scales of the training speakers is avoided. We show that using single Gaussian densities for selecting the normalization scales in training results in lower error rates than using mixture densities. For VTN in recognition, we propose an improvement of the well-known multiple-pass strategy: by using an unnormalized acoustic model for the first recognition pass instead of a normalized model lower error rates are obtained. In recognition tests, this method is compared with a fast variant of VTN. The multiple-pass strategy is an efficient method but it is suboptimal because the normalization scale and the word sequence are determined sequentially. We found that for telephone digit string recognition this suboptimality reduces the VTN gain in recognition performance by 30% relative. On the German spontaneous scheduling task Verbmobil, the WSJ task and the German telephone digit string corpus SieTill the proposed methods for VTN reduce the error rates significantly.

We present an overview of the RWTH Aachen large vocabulary continuous speech recognizer. The recognizer is based on continuous density hidden Markov models and a time-synchronous left-to-right beam search strategy. Experimental results on the ARPA Wall Street Journal (WSJ) corpus verify the effects of several system components, namely linear discriminant analysis, vocal tract normalization, pronunciation lexicon and cross-word triphones, on the recognition performance.

Although speaker normalization is attempted in very different manners, vocal tract normalization (VTN) and speaker adaptive training (SAT) share many common properties. We show that both lead to more compact representations of the phonetically relevant variations of the training data and that both achieve improved error rate performance only if a complementary normalization or adaptation operation is conducted on the test data. Algorithms for fast test speaker enrolment are presented for both normalization methods: in the framework of SAT, a pre-transformation step is proposed, which alone, i.e. without subsequent unsupervised MLLR adaptation, reduces the error rate by almost 10% on the WSJ 5k test sets. For VTN, the use of a Gaussian mixture model makes obsolete a first recognition pass to obtain a preliminary transcription of the test utterance at hardly any loss in performance.

This paper presents a new method for estimating formant frequencies. The formant model is based on a digital resonator. Each resonator represents a segment of the short-time power spectrum. The complete spectrum is modeled by a set of digital resonators connected in parallel. An algorithm based on dynamic programming produces both the model parameters and the segment boundaries that optimally match the spectrum. We used this method in experimental tests that were carried out on the TI digit string data base. The main results of the experimental tests are: (1) the presented approach produces reliable estimates of formant frequencies across a wide range of sounds and speakers; and (2) the estimated formant frequencies were used in a number of variants for recognition. The best set-up resulted in a string error rate of 4.2% on the adult corpus of the TI digit string data base.

In this work we compare two parameter optimization techniques for discriminative training using the MMI criterion: the extended Baum-Welch (EBW) algorithm and the generalized probabilistic descent (GPD) method.Using Gaussian emission densities we found special expressions for the step sizes in GPD, leading to reestimation formula very similar to those derived for the EBW algorithm.Results were produced for both the TI digitstring and the SieTill corpus for continuously spoken American English and German digitstrings.The results for both techniques do not show signi cant dierences.This experimental results support the strong link between EBW and GPD as expected from the analytic comparison.

In this paper we describe experiments with the acoustic front--end of our large vocabulary speech recognition system. In particular, two aspects are studied: 1) linear transforms for feature extraction and 2) the modelling of the emission probabilities. Experiments are reported on a 5000--word task of the ARPA Wall Street Journal database.

This paper presents a new method for estimating formant frequencies. The formant model is based on a digital resonator. Each resonator represents a segment of the short-time power spectrum. The complete spectrum is modeled by a set of digital resonators connected in parallel. An algorithm based on dynamic programming produces both the model parameters and segment boundaries that optimally match the spectrum. The main results of this paper are: (1) modeling formants by digital resonators allows a reliable estimation of formant frequencies; (2) digital resonators can be used efficiently in connection with dynamic programming; and (3) a recognition test with formant frequencies results in a string error rate of 4.8% on the adult corpus of the TI digit string database.

In this paper we investigate Linear Discriminant Analysis (LDA) for the TI connected digit recognition task (TI task) and the Wall Street Journal large vocabulary recognition task (WSJ task). In addition to previous variants of LDA implementations, we avoided the explicit incorporation of derivatives in the acoustic vector. Instead a sliding window without derivatives was used. This large-sized vector was then taken to extract the features by an LDA transformation. Tests for this feature generation were performed both for Laplacian and Gaussian densities. 1. INTRODUCTION It is a well known fact that the performance of a pattern recognition system depends heavily on the type of observations that are used in the system. Several methods are employed in practice which often consist of two stages: First the acoustic signal is transformed from time domain into frequency domain using a Fourier transformation or the like. Second the spectral components of the resulting acoustic vector are th...

In this paper we describe the optimization of 'conventional ' template matching techniques for connected digit recognition (TI/NIST connected digit corpus). In particular we carried out a series of experiments in which we studied various aspects of signal processing, acoustic modeling, mixture densities and linear transforms of the acoustic vector. After all optimization steps, our best string error rate on the TI/NIST connected digit corpus was 1.71% for single densities and 0.74% for mixture densities. 1. INTRODUCTION Over the last five years much progress has been made in connected digit recognition [3, 7, 8, 9]. This paper describes how the systematic optimization of various components of a 'conventional' recognition system leads to high performance comparable with other systems that use much more complicated techniques. Experimental results on the adult corpus of the TI/NIST connected digit corpus are given. The optimization steps presented in this paper are: 1. Several methods f...