KR20010082390A - The Performance Improvement of Speech recognition system using Hierarchical Classification Method and Optimal Candidate Model - Google Patents

Abstract

PURPOSE: A broad classification method of an HMM(Hidden Markov Model) improving speech recognition speed and a method for detecting an ideal candidate are provided to improve a speech recognition speed of a speech recognition system by using a broad classification method. CONSTITUTION: A speech inputted as A-law coding formation is converted a linear PCM. An end point is detected by using an energy. A code book is generated by using five streams such as 12 Mel-Cepstrum, 12 Delta, 12 Acceleration Cepstrum, 3 merged energies(energy+delta energy+delta delta energy), 5 PLP(Perceptual Linear Prediction), and a RASTA cepstrum. The code book is formed by a vector quantization process. The Mel-Cepstrum, Delta-Cepstrum, Acceleration Cepstrum, and PLP generates 256 code book. The merged energy generates 64 code book. The first code index stream passing 12 code books passes an HMM. The state number and the stream number are reduced to improve a speed of the speech recognition system. An N-state HMM for selecting a candidate is located at a front end portion of a full state HMM. The N-state HMM and the full state HMM form 1445 as word number, respectively.

Description

{The Performance Improvement of Speech recognition System Using Hierarchical Classification Method and Optimal Candidate Model}

The speech recognition methods currently studied include Artificial Neural Net (ANN), Dynamic Time Warping (DTW) algorithm using temporal matching, and Hidden Markov Model (HMM), which is known as a probabilistic method. Among them, the method using HMM is known to have the highest recognition rate. The method using the HMM can also be classified into phonemes and words.

As voice recognition technology has been commercialized, it has emerged as a bigger problem than recognition rate. In fact, a single voice recognition server often handles more than 60 lines. This means that 60 voice recognition threads must be processed at the same time, and real-time speed is hard to expect.

The research on the improvement method of speed processing has been mainly conducted on the speech recognizer whose phoneme recognition unit is phoneme. When the recognition unit is phoneme, since the basic structure is in the form of a lexicon tree, the searching time is inevitably shorter than that of the word unit, and there are various studies on related tree searching. Since the word recognition unit recognizer basically needs to perform depth first search and go through the corresponding HMM of all words, it takes a long time. In addition, each HMM model occupies memory when the recognition process is performed. The memory occupancy of the word recognition unit model is much larger than that of the phoneme recognition unit model, and the speed difference that occurs here is also large. However, despite these shortcomings, research on word recognition unit recognizer is necessary. The phoneme recognition unit recognizer is significantly lower than the word recognition unit recognizer in terms of the performance of word speech recognition.

The present invention is directed to speed improvement for a word recognition unit recognizer having such characteristics. The purpose of this study is to establish a speed improvement method while maintaining the advantage of high recognition rate of word recognition unit in word recognition.

The hierarchical large classification technique is applied to the HMM model to improve the speed. Usually, recognition processing time is mostly search time for recognition model. The large classification technique in the present invention is a method of reducing the search space by applying HMM models having different state numbers in multiple stages. In order to reduce the search space, candidates are selected using a small number of state HMM models. This experiment was performed to find the optimal case by adjusting the frequency of candidate inclusion for a few number of state models. For such a small number of state HMMs, the most ideal case is found by flexibly changing the setting of the number of candidates and the number of associated feature parameters.

Figure 1. Concept of Large Classification Technique

The hierarchical large classification technique is applied to the HMM model to improve the speed. Usually, recognition processing time is mostly search time for recognition model. The large classification technique in this paper is to reduce the search space by applying HMM models with different number of states in multiple stages. In order to reduce the search space, candidates are selected using a small number of state HMM models. This experiment was performed to find the optimal case by adjusting the frequency of candidate inclusion for a few number of state models. For such a small number of state HMMs, the setting of the number of candidates and the number of associated feature parameters varies flexibly to find the most ideal case.

Figure 2. Structure of the HMM

HMM is a probabilistic model most often used for speech recognition. Its components consist of factors such as number of states, transitions, and probability of transition. As with normal speech recognition, use a left-to-right model or a similar model that transitions from left to right over time. This structure serves as a reference pattern for the target word, and models the temporal change characteristic. In sentence-independent, the ergodic topology is used as shown in (b) of Figure 2. The model is structured so that all states are completely connected, so you can transition to any state and return to your state. Therefore, although not accurate, each state automatically forms a broad phonetic class by a learning sentence, and the output probability is calculated while transitioning to the phoneme sequences most similar to the input sentence upon recognition. There is an advantage.

Figure 3. Configuration and steps of speech recognition system

The speech recognition step is a structure that selects the largest likelihood ratio result based on speech segment detection, feature parameter extraction, vector quantization, and HMM model.

The structure of speech recognition system is divided into speech segment detection stage, acoustic analysis stage, recognition unit division stage, and recognition stage. After detecting the speech section for recognition using zero crossing rate and algebraic energy, information for recognition is extracted through preprocessing and feature parameter extraction. This information is expressed as several representative vectors through the vector quantization process, and then the one-dimensional code index string, which is its output, is trained using the HMM model. The recognizer of this paper trained 1445 securities stock words using the HMM model.

This system converts voice sound input in A-law coding format into linear PCM and detects the end point using energy. We then use five feature parameters (5 streams). Using the 12th Mel-Cepstrum, the 12th Delta, the 12th Acceleration Cepstrum, and the 3rd Merged Energy (energy + delta energy + delta delta energy) and the 5th Perceptual Linear Prediction (PLP) and RASTA Cepstrum reflecting the auditory signal To generate a codebook.

The codebook is generated through a vector quantization (VQ) process. This compresses the information by representing the words in several representative vectors. The method uses the sum of the distortion distances from the nearest representative vector as the confirmation score, and can also be used in text-dependent mode. Mel-Cepsrum, Delta-Cepsrum, Acceleration Cepstrum, and PLP generate 256-order codebooks, and Merged Energy produces 64-order codebooks. The result is a code word sequence, expressed as a one-dimensional code index.

The primary code index string that passes through the 12th codebook passes through the HMM model. There are two factors that affect speed in speech recognition systems. First, the long time to pass through all 1445 total HMMs is caused. The total number of states HMM has 7 states for two syllables, 15 for 3 and 4 syllables, and 21 for 5 syllables, depending on the length of the syllable. A lot of time is spent passing through this HMM model with the overall state number. The processing time of the HMM model is proportional to the number of states. Second, the number of feature parameters also affects the speed of the system. The number of feature parameters used in the present invention is five, and the speed when using two and using five is very different.

To overcome this problem, a small number of N-state HMM models for selecting candidates are placed in front of the full-state HMM. Not only does it reduce the number of states but also the number of feature parameters, so it does not take long processing time. Only elected candidates pass the full-state HMM. To this end, a small number of N-state HMMs and full-state HMMs form 1445 words, respectively.

In a small number of state HMMs to find candidates, we need to find a state number N that finds a suitable candidate while minimizing time. First, the number of states is defined as 1, 2, 3, 4, and 5 as candidates for N. In consideration of the fact that the two syllable words in the total number of HMMs had seven number of states, the maximum number of N was tested up to 5. The candidate extraction accuracy is shown in Figures 4 and 5. As a result, the three-state HMM model is the most efficient. Candidate extraction accuracy does not show a big difference between the three, four, and five states, but there are many differences in speed. State counts 1 and 2 are difficult to play the role of candidate extraction.

The index sequence first passes 1445 tri-state HMMs to obtain the likelihood ratios. N candidates are extracted from the obtained likelihood ratio values. The extracted N candidates recognize the maximum likelihood ratio value by applying the full state HMM as the corresponding word. When the candidate is extracted, the number of feature parameters applied is classified into various cases and tested. Experiments were performed to obtain the most optimized results by varying the number of candidates and the type and number of feature parameters applied. The number of candidates could be classified into one or two stages.

The present invention is a large class multi-stage recognizer in consideration of feature parameters and state numbers, and has an excellent effect on speed improvement.

Usually, a system that prefers recognition rate uses a word recognition unit recognizer. However, due to the speed problem, it was not used in a system such as a multiple access environment. The present invention has the meaning that solved the speed problem in a stable word recognition unit.

Claims (2)

In the speech recognition process, a large classification technique with variations in the number of states and feature parameters of the HMM is used. It is applied to multiple access dial-up service and embedded system where speed problem of large amount of computation is important.