In this paper, an MLLR-like adaptation approach is proposed whereby the transformation of the means is performed deterministically based on linearization of VTLN. Biases and adaptation of the variances are estimated statistically by the EM algorithm. In the discrete frequency domain, we show that under certain approximations, frequency warping with Mel-£lterbank-based MFCCs equals a linear tran...

Vocal Tract Length Normalization (VTLN) and Maximum Likelihood Linear Regression (MLLR) are two approaches to reduce the degradation in speech recognition performance caused by variation of speakers. This paper derives a novel efficient adaptation algorithm from the two techniques. Based on prior knowledge of usual VTLN, an approximate constrained-form linear transformation is obtained. The tra...

In this paper, we considered two representative speaker adaptation (SA) approaches – MAP and MLLR techni -ques – for our Korean isolated word recognition task. In addition, we proposed a new speaker adaptation alg orithm to improve the performance of MAP technique. It is based on least squares method between MAP ada pted mean vectors and the corresponding SI mean vect ors. The results of our ex...

With the tendency of posterior probability taken into account, a state-restructuring method is proposed based on confusions between HMM states. In the method, HMM state is restructured by sharing Gaussian components with its related states and the re-estimation of the increased-parameters, i.e., the inter-state weights, is derived under the EM framework. Experiments are performed on speaker-ind...

In this paper, we expand on a previously proposed algorithm entitled Structural Maximum Likelihood Eigenspace Mapping (SMLEM) [5, 6] for rapid speaker adaptation by exploring a variety of model clustering methods and incorporating a multi-stream approach. The SMLEM algorithm directly adapts speaker independent acoustic models to a test speaker by mapping the mixture Gaussian components from a s...

Transformation based speaker adaptation techniques, such as Maximum Likelihood Linear Regression (MLLR) [1] require a large amount of adaptation data to robustly estimate the transform matrices. In this paper, we present a new adaptation scheme that adjusts the adaptation data according to the feedback from recognizer. By giving different weights to different parts of the adaptation data, the p...

We present a simplified derivation of the extended Baum-Welch procedure, which shows that it can be used for Maximum Mutual Information (MMI) of a large class of continuous emission density hidden Markov models (HMMs). We use the extended Baum-Welch procedure for discriminative estimation of MLLR-type speaker adaptation transformations. The resulting adaptation procedure, termed Conditional Max...

Transformation based speaker adaptation techniques, such as Maximum Likelihood Linear Regression (MLLR) require a large amount of adaptation data to robustly estimate the transform matrices. In this paper, we present a new adaptation scheme to make use of adaptation data more effectively, which adjusts the adaptation data according to the decoding results on the same adaptation set. The adjustm...

Adaptation techniques are necessary in automatic speech recognizers to improve a recognition accuracy. Linear Transformation methods (MLLR or fMLLR) are the most favorite in the case of limited available data. The fMLLR is the feature-space transformation. This is the advantage with contrast to MLLR that transforms the entire acoustic model. The classical fMLLR estimation involves maximization ...

This paper evaluates the recognition performance of a system using acoustic models transformed across language boundaries. Parameters of hidden Markov models (HMMs) trained on speaker independent English data are adapted using Afrikaans adaptation data to realise speaker dependent, multispeaker and speaker independent Afrikaans models. Adaptation is performed using maximum a posteriori probabil...