• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.37-41
• /
• 2020

본 논문에서는 Joint CTC/Attention 모델에 CTC ratio scheduling을 이용한 end-to-end 한국어 음성인식을 연구하였다. Joint CTC/Attention은 CTC와 attention의 장점을 결합한 모델로서 attention, CTC 단일 모델보다 좋은 성능을 보여주지만, 학습이 진행될수록 CTC가 attention의 학습을 저해하는 요인이 된다. 본 논문에서는 이러한 문제를 해결하기 위해, 학습 진행에 따라 CTC의 비율(ratio)를 줄여나가는 CTC ratio scheduling 방법을 제안한다. CTC ratio scheduling를 이용하여 학습한 결과물은 기존 Joint CTC/Attention, 단일 attention 모델 대비 좋은 성능을 보여주는 것을 확인하였다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.6 no.11
• /
• pp.2800-2812
• /
• 2012

The purpose of this study was to examine the syllable-word boundary misalignment cost on word segmentation in Korean continuous speech. Previous studies have demonstrated the important role of syllabification in speech segmentation. The current study investigated whether the resyllabification process affects word recognition in Korean continuous speech. In Experiment I, under the misalignment condition, participants were presented with stimuli in which a word-final consonant became the onset of the next syllable. (e.g., /k/ in belsak ingan becomes the onset of the first syllable of ingan 'human'). In the alignment condition, they heard stimuli in which a word-final vowel was also the final segment of the syllable (e.g., /eo/ in heulmeo ingan is the end of both the syllable and word). The results showed that word recognition was faster and more accurate in the alignment condition. Experiment II aimed to confirm that the results of Experiment I were attributable to the resyllabification process, by comparing only the target words from each condition. The results of Experiment II supported the findings of Experiment I. Therefore, based on the current study, we confirmed that Korean, a syllable-timed language, has a misalignment cost of resyllabification.

• MALSORI
• /
• no.65
• /
• pp.153-168
• /
• 2008

We propose a new method to apply shape analysis techniques to a computational auditory scene analysis (CASA)-based speech separation system. The conventional CASA-based speech separation system extracts speech signals from a mixture of speech and noise signals. In the proposed method, we complement the missing speech signals by applying the shape analysis techniques such as labelling and distance function. In the speech separation experiment, the proposed method improves signal-to-noise ratio by 6.6 dB. When the proposed method is used as a front-end of speech recognizers, it improves recognition accuracy by 22% for the speech-shaped stationary noise condition and 7.2% for the two-talker noise condition at the target-to-masker ratio than or equal to -3 dB.

• Speech Sciences
• /
• v.10 no.4
• /
• pp.117-124
• /
• 2003

Speaker recognition is the technology that confirms the identification of speaker by using the characteristic of speech. Such technique is classified into speaker identification and speaker verification: The first method discriminates the speaker from the preregistered group and recognize the word, the second verifies the speaker who claims the identification. This method that extracts the information of speaker from the speech and confirms the individual identification becomes one of the most efficient technology as the service via telephone network is popularized. Some problems, however, must be solved for the real application as follows; The first thing is concerning that the safe method is necessary to reject the imposter because the recognition is not performed for the only preregistered customer. The second thing is about the fact that the characteristic of speech is changed as time goes by, So this fact causes the severe degradation of recognition rate and the inconvenience of users as the number of times to utter the text increases. The last thing is relating to the fact that the common characteristic among speakers causes the wrong recognition result. The silence parts being included the center of speech cause that identification rate is decreased. In this paper, to make improvement, We proposed identification rate can be improved by removing silence part before processing identification algorithm. The methods detecting speech area are zero crossing rate, energy of signal detect end point and starting point of the speech and process DTW algorithm by using two methods in this paper. As a result, the proposed method is obtained about 3% of improved recognition rate compare with the conventional methods.

• MALSORI
• /
• no.65
• /
• pp.93-103
• /
• 2008

In this paper, we propose a method for optimizing a multiple pronunciation dictionary used for modeling pronunciation variations of non-native speech. The proposed method removes some confusable pronunciation variants in the dictionary, resulting in a reduced dictionary size and less decoding time for automatic speech recognition (ASR). To this end, a confusability measure is first defined based on the Levenshtein distance between two different pronunciation variants. Then, the number of phonemes for each pronunciation variant is incorporated into the confusability measure to compensate for ASR errors due to words of a shorter length. We investigate the effect of the proposed method on ASR performance, where Korean is selected as the target language and Korean utterances spoken by Chinese native speakers are considered as non-native speech. It is shown from the experiments that an ASR system using the multiple pronunciation dictionary optimized by the proposed method can provide a relative average word error rate reduction of 6.25%, with 11.67% less ASR decoding time, as compared with that using a multiple pronunciation dictionary without the optimization.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.17 no.12
• /
• pp.2793-2799
• /
• 2013

This paper proposes an effective feature compensation method to improve speech recognition performance in time-varying background noise condition. The proposed method employs principal component analysis to improve the variational model composition method. The proposed method is employed to generate multiple environmental models for the PCGMM-based feature compensation scheme. Experimental results prove that the proposed scheme is more effective at improving speech recognition accuracy in various SNR conditions of background music, compared to the conventional front-end methods. It shows 12.14% of average relative improvement in WER compared to the previous variational model composition method.

• MALSORI
• /
• no.64
• /
• pp.105-119
• /
• 2007

Noise reduction technologies are indispensable to achieve acceptable speech quality in VoIP systems. This paper proposes a Wiener filter optimized to the estimated SNR of noisy speech for the noise reduction in VoIP environments. The proposed noise canceller is applied as a pre-processor before speech encoding. The performance of the proposed method is evaluated by the PESQ in various noisy conditions. In this paper, the proposed algorithm is applied to G.711, G.723.1, and G.729A which are all VoIP speech codecs. The PESQ results show that the performance of our proposed noise reduction scheme outperforms those of the noise suppression in the IS-127 EVRC and the ETSI standard for the advanced distributed speech recognition front-end.

• The Journal of the Acoustical Society of Korea
• /
• v.16 no.2
• /
• pp.41-48
• /
• 1997

This paper proposes the methods for improving the recognition rate of theARS, especially equipped with the speech recognition capability. Telephone speech, which is the input to the ARS, is usually affected by the announcements from the system, channel noise, and channel distortion, thus directly applying the recognition algorithm developed for clean speech to the noisy telephone speech will bring the significant performance degradation. To cope with this problem, this paper proposes three methods: 1)the accurate detection of the inputting instant of the speech in order to immediately turn off the announcements from the system at that instant, 2)the effective end-point detection of the noisy telephone speech on the basis of Teager energy, and 3)the SDCN-based compensation of the channel distortion. Experiments on speaker-independent, noisy telephone speech reveal that the combination of the above three proposed methods provides great improvements on the recognition rate over the conventional method, showing about 77% in contrast to only 23%.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.5 no.3
• /
• pp.451-458
• /
• 2001

In this paper, we propose speech control system for a various control device in the car with real time control speech. A real time speech control system is detected start-end points from speech data processing by A/D conversion, and recognize by one pass dynamic programming method. The results displays a monitor, and transports control data to control interfaces. The HMM model is modeled by a continuous control speech consists of control speech and digit speech for controlling of a various control device in the car The recognition rates is an average 97.3% in case of word & control speech, and is an average 96.3% in case of digit speech.

• Speech Sciences
• /
• v.8 no.2
• /
• pp.83-90
• /
• 2001

We proposed a probabilistic approach of using syllable bigrams to the word-spacing problem. Syllable bigrams are extracted and the frequencies are calculated for the large corpus of 12 million words. Based on the syllable bigrams, we performed three experiments: (1) automatic word-spacing, (2) detection and correction of word-spacing errors for spelling checker, and (3) automatic insertion of a space at the end of line in the character recognition system. Experimental results show that the accuracy ratios are 97.7 percent, 82.1 percent, and 90.5%, respectively.