• ETRI Journal
• /
• v.18 no.1
• /
• pp.29-41
• /
• 1996

This paper describes a method to select a suitable feature for speech recognition using information theoretic measure. Conventional speech recognition systems heuristically choose a portion of frequency components, cepstrum, mel-cepstrum, energy, and their time differences of speech waveforms as their speech features. However, these systems never have good performance if the selected features are not suitable for speech recognition. Since the recognition rate is the only performance measure of speech recognition system, it is hard to judge how suitable the selected feature is. To solve this problem, it is essential to analyze the feature itself, and measure how good the feature itself is. Good speech features should contain all of the class-related information and as small amount of the class-irrelevant variation as possible. In this paper, we suggest a method to measure the class-related information and the amount of the class-irrelevant variation based on the Shannon's information theory. Using this method, we compare the mel-scaled FFT, cepstrum, mel-cepstrum, and wavelet features of the TIMIT speech data. The result shows that, among these features, the mel-scaled FFT is the best feature for speech recognition based on the proposed measure.

• Phonetics and Speech Sciences
• /
• v.12 no.4
• /
• pp.63-71
• /
• 2020

Speech recognition is one of the areas actively commercialized using deep learning and machine learning techniques. However, the majority of speech recognition systems on the market are developed on data with limited diversity of speakers and tend to perform well on typical adult speakers only. This is because most of the speech recognition models are generally learned using a speech database obtained from adult males and females. This tends to cause problems in recognizing the speech of the elderly, children and people with dialects well. To solve these problems, it may be necessary to retain big database or to collect a data for applying a speaker adaptation. However, this paper proposes that a new end-to-end speech recognition method consists of an acoustic augmented recurrent encoder and a transformer decoder with linguistic prediction. The proposed method can bring about the reliable performance of acoustic and language models in limited data conditions. The proposed method was evaluated to recognize Korean elderly and children speech with limited amount of training data and showed the better performance compared of a conventional method.

• International Journal of Fuzzy Logic and Intelligent Systems
• /
• v.4 no.2
• /
• pp.149-154
• /
• 2004

In this research, we selected the speech recognition to implement the electric wheelchair system as a method to control it by only using the speech and used DTW (Dynamic Time Warping), which is speaker-dependent and has a relatively high recognition rate among the speech recognitions. However, it has to have small memory and fast process speed performance under consideration of real-time. Thus, we introduced VQ (Vector Quantization) which is widely used as a compression algorithm of speaker-independent recognition, to secure fast recognition and small memory. However, we found that the recognition rate decreased after using VQ. To improve the recognition rate, we applied ART2 (Adaptive Reason Theory 2) algorithm as a post-process algorithm to obtain about 5% recognition rate improvement. To utilize ART2, we have to apply an error range. In case that the subtraction of the first distance from the second distance for each distance obtained to apply DTW is 20 or more, the error range is applied. Likewise, ART2 was applied and we could obtain fast process and high recognition rate. Moreover, since this system is a moving object, the system should be implemented as an embedded one. Thus, we selected TMS320C32 chip, which can process significantly many calculations relatively fast, to implement the embedded system. Considering that the memory is speech, we used 128kbyte-RAM and 64kbyte ROM to save large amount of data. In case of speech input, we used 16-bit stereo audio codec, securing relatively accurate data through high resolution capacity.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.23 no.1
• /
• pp.33-38
• /
• 2019

This paper proposes an effective recognition method of VPI patient's speech employing DNN-HMM-based speech recognition system, and evaluates the recognition performance compared to GMM-HMM-based system. The proposed method employs speaker adaptation technique to improve VPI speech recognition. This paper proposes to use simulated VPI speech for generating a prior model for speaker adaptation and selective learning of weight matrices of DNN, in order to effectively utilize the small size of VPI speech for model adaptation. We also apply Linear Input Network (LIN) based model adaptation technique for the DNN model. The proposed speaker adaptation method brings 2.35% improvement in average accuracy compared to GMM-HMM based ASR system. The experimental results demonstrate that the proposed DNN-HMM-based speech recognition system is effective for VPI speech with small-sized speech data, compared to conventional GMM-HMM system.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.3
• /
• pp.10-16
• /
• 1999

In this paper, we have studied Continuous Speech Recognition System of Speaker Adaptation using MAPE (Maximum A Posteriori Probability Estimation) which can adapt any small amount of adaptation speech data. Speaker adaptation is performed by the method of MAPB after Concatenation training which is making sentence unit HMM linked by syllable unit HMM and Viterbi segmentation classifies speech data to be adaptation into segmentation of syllable unit data automatically without hand labelling. For car control speech the recognition rates of adaptation of HMM was 77.18% which is approximately 6% improvement over that of unadapted HMM.(in case of O(n)DP)

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.3
• /
• pp.56-62
• /
• 2001

In this paper, we adopt a missing data theory to speech recognition. It can be used in order to maintain high performance of speech recognizer when the missing data occurs. In general, hidden Markov model (HMM) is used as a stochastic classifier for speech recognition task. Acoustic events are represented by continuous probability density function in continuous density HMM(CDHMM). The missing data theory has an advantage that can be easily applicable to this CDHMM. A marginalization method is used for processing missing data because it has small complexity and is easy to apply to automatic speech recognition (ASR). Also, a spectral subtraction is used for detecting missing data. If the difference between the energy of speech and that of background noise is below given threshold value, we determine that missing has occurred. We propose a new method that examines the reliability of detected missing data using voicing probability. The voicing probability is used to find voiced frames. It is used to process the missing data in voiced region that has more redundant information than consonants. The experimental results showed that our method improves performance than baseline system that uses spectral subtraction method only. In 452 words isolated word recognition experiment, the proposed method using the voicing probability reduced the average word error rate by 12％ in a typical noise situation.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.6
• /
• pp.655-661
• /
• 2022

Since speech impairment is prevalent in patients with Parkinson's disease (PD), speech recognition systems suitable for these patients are needed. In this paper, we propose a speech recognition system that effectively recognizes the speech of patients with PD. The speech recognition system is firstly pre-trained with the Globalformer using the speech data from healthy people, and then fine-tuned using relatively small amount of speech data from the patient with PD. For this analysis, we used the speech dataset of healthy people built by AI hub and that of patients with PD collected at Inha University Hospital. As a result of the experiment, the proposed speech recognition system recognized the speech of patients with PD with Character Error Rate (CER) of 22.15 %, which was a better result compared to other methods.

• Speech Sciences
• /
• v.11 no.3
• /
• pp.143-156
• /
• 2004

In this paper, we implemented a vocabulary-independent speech recognizer using the TMS320VC33 DSP. For this implementation, we had developed very small-sized recognition engine based on diphone sub-word unit, which is especially suited for embedded applications where the system resources are severely limited. The recognition accuracy of the developed recognizer with 1 mixture per state and 4 states per diphone is 94.5% when tested on frequently-used 2000 words set. The design of the hardware was focused on minimal use of parts, which results in reduced material cost. The finally developed hardware only includes a DSP, 512 Kword flash ROM and a voice codec. In porting the recognition engine to the DSP, we introduced several methods of using data and program memory efficiently and developed the versatile software protocol for host interface. Finally, we also made an evaluation board for testing the developed hardware recognition module.

• 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.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.83-88
• /
• 2023

This study investigates the fine-tuning of large-scale Automatic Speech Recognition (ASR) models, specifically OpenAI's Whisper model, for domain-specific applications using the KsponSpeech dataset. The primary research questions address the effectiveness of targeted lexical item emphasis during fine-tuning, its impact on domain-specific performance, and whether the fine-tuned model can maintain generalization capabilities across different languages and environments. Experiments were conducted using two fine-tuning datasets: Set A, a small subset emphasizing specific lexical items, and Set B, consisting of the entire KsponSpeech dataset. Results showed that fine-tuning with targeted lexical items increased recognition accuracy and improved domain-specific performance, with generalization capabilities maintained when fine-tuned with a smaller dataset. For noisier environments, a trade-off between specificity and generalization capabilities was observed. This study highlights the potential of fine-tuning using minimal domain-specific data to achieve satisfactory results, emphasizing the importance of balancing specialization and generalization for ASR models. Future research could explore different fine-tuning strategies and novel technologies such as prompting to further enhance large-scale ASR models' domain-specific performance.