• MALSORI
• /
• no.62
• /
• pp.113-131
• /
• 2007

It is reported that the segmental feature based recognition system shows better results than conventional feature based system in the previous studies. On the other hand, the various studies of combining neural network and hidden Markov models within a single system are done with expectations that it may potentially combine the advantages of both systems. With the influence of these studies, tandem approach was presented to use neural network as the classifier and hidden Markov models as the decoder. In this paper, we applied the trend information of segmental features to tandem architecture and used posterior probabilities, which are the output of neural network, as inputs of recognition system. The experiments are performed on Auroral database to examine the potentiality of the trend feature based tandem architecture. From the results, the proposed system outperforms on very low SNR environments. Consequently, we argue that the trend information on tandem architecture can be additionally used for traditional MFCC features.

• Proceedings of the KSPS conference
• /
• 2002.11a
• /
• pp.205-208
• /
• 2002

Bimodal speech recognition based on lip reading has been studied as a representative method of speech recognition under noisy environments. There are three integration methods of speech and lip modalities as like direct identification, separate identification and dominant recording. In this paper we evaluate the robustness of lip reading methods under the assumption that lip parameters are estimated with errors. We show that the dominant recording approach is more robust than other methods with lip reading experiments. Also, a measure of lip parameter degradation is proposed. This measure can be used in the determination of weighting values of video information.

• MALSORI
• /
• v.55
• /
• pp.45-60
• /
• 2005

The purpose of this paper is to represent values of acoustic cues for Korean oral stops in the multi-dimensional space, and to attempt to find possible relationships among acoustic cues through correlation analyses. The acoustic cues used for differentiation of 3 types of Korean stops are closure duration, voice onset time and fundamental frequency of a vowel after a stop. The values of these cues are plotted in the two and three dimensional space to see what the critical cues are for separation of different types of stops. Correlation coefficient analyses show that multi-variate approach to statistical analysis is legitimate, and that there are statistically significant relationships among acoustic cues but Oey are not strong enough to make the conjecture that there is a possible relationship among the articulatory or laryngeal mechanisms employed by the acoustic cues.

• MALSORI
• /
• no.56
• /
• pp.195-206
• /
• 2005

Discrimination between speech and music is important in many multimedia applications. In this paper, focusing on the spectral change characteristics of speech and music, we propose a new method of speech/music discrimination based on cepstral distance. Instead of using cepstral distance between the frames with fixed interval, the minimum of cepstral distances among neighbor frames is employed to increase discriminability between fast changing music and speech. And, to prevent misclassification of speech segments including short pause into music, short pause segments are excluded from computing cepstral distance. The experimental results show that proposed method yields the error rate reduction of$68\%$, in comparison with the conventional approach using cepstral distance.

• MALSORI
• /
• no.56
• /
• pp.207-223
• /
• 2005

The goal of language model adaptation is to improve the background language model with a relatively small adaptation corpus. This study presents a language model adaptation technique where additional text data for the adaptation do not exist. We propose the information retrieval (IR) technique with N-gram language modeling to collect the adaptation corpus from baseline text data. We also propose to use a dynamic language model interpolation coefficient to combine the background language model and the adapted language model. The interpolation coefficient is estimated from the word hypotheses obtained by segmenting the input speech data reserved for held-out validation data. This allows the final adapted model to improve the performance of the background model consistently The proposed approach reduces the word error rate by $13.6\%$ relative to baseline 4-gram for two-hour broadcast news speech recognition.

• MALSORI
• /
• no.56
• /
• pp.225-235
• /
• 2005

Speech unit concatenation with a large database is presently the most popular method for speech synthesis. In this approach, the mismatches at the unit boundaries are unavoidable and become one of the reasons for quality degradation. This paper proposes an algorithm to reduce undesired discontinuities between the subsequent units. Optimal matching points are calculated in two steps. Firstly, the fullback-Leibler distance measurement is utilized for the spectral matching, then the unit sliding and the overlap windowing are used for the waveform matching. The proposed algorithm is implemented for the corpus-based unit concatenating Korean text-to-speech system that has an automatically labeled database. Experimental results show that our algorithm is fairly better than the raw concatenation or the overlap smoothing method.

• MALSORI
• /
• no.44
• /
• pp.83-92
• /
• 2002

Speech recognition performance is severly degraded under noisy envrionments. One approach to cope with this problem is audio-visual speech recognition. In this paper, we discuss the experiment results of bimodal speech recongition based on enhanced speech feature vectors using lip information. We try various kinds of speech features as like linear predicion coefficient, cepstrum, log area ratio and etc for transforming lip information into speech parameters. The experimental results show that the cepstrum parameter is the best feature in the point of reconition rate. Also, we present the desirable weighting values of audio and visual informations depending on signal-to-noiso ratio.

• MALSORI
• /
• no.49
• /
• pp.63-75
• /
• 2004

Most existing telephone speech transmitted in current public networks is band-limited to 0.3-3.4 kHz. Compared with wideband speech(0-8 kHz), the narrowband speech lacks low-band (0-0.3 kHz) and high-band(3.4-8 kHz) components of sound. As a result, the speech is characterized by the reduced intelligibility and a muffled quality, and degraded speaker identification. Bandwidth extension is a technique to provide wideband speech quality, which means reconstruction of low-band and high-band components without any additional transmitted information. Our new approach considers to exploit harmonic synthesis method for reconstruction of low-band speech over the CELP coded speech. A spectral distortion measurement and listening test are introduced to assess the proposed method, and the improvement of synthesized speech quality was verified.

• MALSORI
• /
• no.49
• /
• pp.159-174
• /
• 2004

The development of the durational component of prosody modelling was overviewed and discussed in text-to-speech conversion of spoken English and Korean, showing the strengths and weaknesses of each approach. The possibility of integrating linguistic feature effects into the duration modelling of TTS systems was also investigated. This paper claims that current approaches to language timing synthesis still require an understanding of how segmental duration is affected by context. Three modelling approaches were discussed: sequential rule systems, Classification and Regression Tree (CART) models and Sums-of-Products (SoP) models. The CART and SoP models show good performance results in predicting segment duration in English, while it is not the case in the SoP modelling of spoken Korean.

• MALSORI
• /
• no.48
• /
• pp.155-173
• /
• 2003

This paper proposes a noise reduction-based speech detection method under telephone channel environments. We adopt the AURORA front-end noise reduction algorithm based on the two-stage mel-warped Wiener filter approach as a preprocessor for the frequency domain speech detector. The speech detector utilizes mel filter-bank based useful band energies as its feature parameters. The preprocessor firstly removes the adverse noise components on the incoming noisy speech signals and the speech detector at the next stage detects proper speech regions for the noise-reduced speech signals. Experimental results show that the proposed noise reduction-based speech detection method is very effective in improving not only the performance of the speech detector but also that of the subsequent speech recognizer.