• MALSORI
• /
• no.58
• /
• pp.35-44
• /
• 2006

Maximum likelihood linear regression (MLLR) adaptation experiences severe performance degradation with very tiny amount of adaptation data. Eigenspace- based MLLR, as an alternative to MLLR for fast speaker adaptation, also has a weak point that it cannot deal with the mismatch between training and testing environments. In this paper, we propose a simultaneous fast speaker and environment adaptation based on eigenspace-based MLLR. We also extend the sub-stream based eigenspace-based MLLR to generalize the eigenspace-based MLLR with bias compensation. A vocabulary-independent word recognition experiment shows the proposed algorithm is superior to eigenspace-based MLLR regardless of the amount of adaptation data in diverse noisy environments. Especially, proposed sub-stream eigenspace-based MLLR with bias compensation yields 67% relative improvement with 10 adaptation words in 10 dB SNR environment, in comparison with the conventional eigenspace-based MLLR.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.6
• /
• pp.452-456
• /
• 2003

We discuss how to reduce the number of inverse matrix and its dimensions requested in MLLR framework for speaker adaptation. To find a smaller set of variables with less redundancy, we adapt PCA (principal component analysis) and ICA (independent component analysis) that would give as good a representation as possible. The amount of additional computation when PCA or ICA is applied is as small as it can be disregarded. 10 components for ICA and 12 components for PCA represent similar performance with 36 components for ordinary MLLR framework. If dimension of SI model parameter is n, the amount of computation of inverse matrix in MLLR is proportioned to O(n⁴). So, compared with ordinary MLLR, the amount of total computation requested in speaker adaptation is reduced by about 1/81 in MLLR with PCA and 1/167 in MLLR with ICA.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.29 no.1C
• /
• pp.65-71
• /
• 2004

We improved the MLLR speaker adaptation algorithm with reduction of the order of HMM parameters using PCA(Principle Component Analysis) or ICA(Independent Component Analysis). To find a smaller set of variables with less redundancy, we adapt PCA(principal component analysis) and ICA(independent component analysis) that would give as good a representation as possible, minimize the correlations between data elements, and remove the axis with less covariance or higher-order statistical independencies. Ordinary MLLR algorithm needs more than 30 seconds adaptation data to represent higher word recognition rate of SD(Speaker Dependent) models than of SI(Speaker Independent) models, whereas proposed algorithm needs just more than 10 seconds adaptation data. 10 components for ICA and PCA represent similar performance with 36 components for ordinary MLLR framework. So, compared with ordinary MLLR algorithm, the amount of total computation requested in speaker adaptation is reduced by about 1/167 in proposed MLLR algorithm.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.7
• /
• pp.539-544
• /
• 2003

This paper describes how to reduce the effect of an occupation threshold by that the transform of mixture components of HMM parameters is controlled in hierarchical tree structure to prevent from over-adaptation. To reduce correlations between data elements and to remove elements with less variance, we employ PCA (Principal component analysis) and ICA (independent component analysis) that would give as good a representation as possible, and decline the effect of over-adaptation. When we set lower occupation threshold and increase the number of transformation function, ordinary MLLR adaptation algorithm represents lower recognition rate than SI models, whereas the proposed MLLR adaptation algorithm represents the improvement of over 2% for the word recognition rate as compared to performance of SI models.

• Phonetics and Speech Sciences
• /
• v.1 no.4
• /
• pp.119-125
• /
• 2009

This paper proposes fast speaker adaptation method using artificially distorted speech in telematics terminal under the car noise environment based on eigenspace-based maximum likelihood linear regression (ES-MLLR). The artificially distorted speech is built from adding the various car noise signals collected from a driving car to the speech signal collected from an idling car. Then, in every environment, the transformation matrix is estimated by ES-MLLR using the artificially distorted speech corresponding to the specific noise environment. In test mode, an online model is built by weighted sum of the environment transformation matrices depending on the driving condition. In 3k-word recognition task in the telematics terminal, we achieve a performance superior to ES-MLLR even using the adaptation data collected from the driving condition.

• Proceedings of the KSPS conference
• /
• 2003.05a
• /
• pp.123-126
• /
• 2003

We discuss how to reduce the number of inverse matrix and its dimensions requested in MLLR framework for speaker adaptation. To find a smaller set of variables with less redundancy, we employ PCA(principal component analysis) and ICA(independent component analysis) that would give as good a representation as possible. The amount of additional computation when PCA or ICA is applied is as small as it can be disregarded. The dimension of HMM parameters is reduced to about 1/3 ~ 2/7 dimensions of SI(speaker independent) model parameter with which speech recognition system represents word recognition rate as much as ordinary MLLR framework. If dimension of SI model parameter is n, the amount of computation of inverse matrix in MLLR is proportioned to O($n^4$). So, compared with ordinary MLLR, the amount of total computation requested in speaker adaptation is reduced to about 1/80~1/150.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.38C no.2
• /
• pp.213-218
• /
• 2013

One of the most popular approaches to parameter adaptation in hidden Markov model (HMM) based systems is the maximum likelihood linear regression (MLLR) technique. In our previous study, we proposed factored MLLR (FMLLR) where each MLLR parameter is defined as a function of a control vector. We presented a method to train the FMLLR parameters based on a general framework of the expectation-maximization (EM) algorithm. Using the proposed algorithm, supplementary information which cannot be included in the models is effectively reflected in the adaptation process. In this paper, we apply the FMLLR algorithm to a pitch sequence as well as spectrum parameters. In a series of experiments on artificial generation of expressive speech, we evaluate the performance of the FMLLR technique and also compare with other approaches to parameter adaptation in HMM-based speech synthesis.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.8 no.4
• /
• pp.219-225
• /
• 2013

This paper presents a method using Maximum-Likelihood Linear Regression (MLLR) adaptation to improve recognition performance of Limabeam algorithm for speech recognition using microphone array. From our investigation on Limabeam algorithm, we can see that the performance of filtering optimization depends strongly on the supporting optimal state sequence and this sequence is created by using Viterbi algorithm trained with HMM model. So we propose an approach using MLLR adaptation for the recognition of speech uttered in a new environment to obtain better optimal state sequence that support for the filtering parameters' optimal step. Experimental results show that the system embedded with MLLR adaptation presents the word correct recognition rate 2% higher than that of original calibrate Limabeam and also present 7% higher than that of Delay and Sum algorithm. The best recognition accuracy of 89.4% is obtained when we use 4 microphones with 5 utterances for adaptation.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.2
• /
• pp.190-198
• /
• 2002

In general, we have used the Viterbi algorithm of Speech recognition for decoding. But a decoder in speaker verification has to recognize same word of every speaker differently. In this paper, we propose a noble decoding algorithm that could replace the typical Viterbi algorithm for the speaker verification system. We utilize for the proposed algorithm the speaker adaptation algorithms that transform feature vectors into the region of the client' characteristics in the speech recognition. There are many adaptation algorithms, but we take MLLR (Maximum Likelihood Linear Regression) and MAP (Maximum A-Posterior) adaptation algorithms for proposed algorithm. We could achieve improvement of performance about 30% of EER (Equal Error Rate) using proposed algorithm instead of the typical Viterbi algorithm.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.8
• /
• pp.727-739
• /
• 2003

In this paper, we propose a generation method of regression classes for adaptation in the HM-Net (Hidden Markov Network) system. The MLLR (Maximum Likelihood Linear Regression) adaptation approach is applied to the HM-Net speech recognition system for expressing the characteristics of speaker effectively and the use of HM-Net in various tasks. For the state level sharing, the context domain state splitting of PDT-SSS (Phonetic Decision Tree-based Successive State Splitting) algorithm, which has the contextual and time domain clustering, is adopted. In each state of contextual domain, the desired phoneme classes are determined by splitting the context information (classes) including target speaker's speech data. The number of adaptation parameters, such as means and variances, is autonomously controlled by contextual domain state splitting of PDT-SSS, depending on the context information and the amount of adaptation utterances from a new speaker. The experiments are performed to verify the effectiveness of the proposed method on the KLE (The center for Korean Language Engineering) 452 data and YNU (Yeungnam Dniv) 200 data. The experimental results show that the accuracies of phone, word, and sentence recognition system increased by 34∼37%, 9%, and 20%, respectively, Compared with performance according to the length of adaptation utterances, the performance are also significantly improved even in short adaptation utterances. Therefore, we can argue that the proposed regression class method is well applied to HM-Net speech recognition system employing MLLR speaker adaptation.