CN106601234A - Implementation method of placename speech modeling system for goods sorting - Google Patents

Abstract

The invention discloses a method for realizing a place name voice modeling system oriented to goods sorting, which includes the following steps: 1) preprocessing the voice signal, including pre-emphasis and framing operations; 2) preprocessing in step 1) 3) check the effect of the voice signal after cutting in step 2), if the cutting is correct, save it in the specified folder; 4) when step 3) After the recording of all voice signals is completed, a hidden Markov model of the voice signal is established, and the established model data is saved. Aiming at the short feature of the place name voice signal, the present invention regards each place name voice signal as a recognition object, directly takes the whole of each place name as an object to establish a hidden Markov model, and the calculation is simple and efficient.

Description

An implementation method of a place name phonetic modeling system oriented to cargo sorting

technical field

The invention relates to the fields of signal processing, pattern recognition and human-computer interaction, in particular to an implementation method of a place name voice modeling system oriented to cargo sorting.

Background technique

At present, in the sorting link of the logistics site, the main sorting method is to confirm the slot where the goods are to be sorted by pressing the buttons. The operator must use the buttons to input the sorting information of the goods, and the operation is time-consuming and troublesome. Using the voice recognition system of place names allows the operator to directly communicate with the sorting system and inform the sorting system of the distribution information of the goods. This method makes the sorting of goods more efficient, fast and time-saving. The voice modeling system of place names can The hidden Markov model is established for the signal to facilitate the speech recognition of place names.

Contents of the invention

The object of the present invention is to provide a method for implementing a voice modeling system for place names oriented to cargo sorting, which is easy to operate and capable of human-computer interaction, in view of the above-mentioned deficiencies in the prior art.

The purpose of the present invention can be achieved through the following technical solutions:

A method for realizing a place name voice modeling system oriented to cargo sorting, said method comprising the following steps:

1) Preprocessing the voice signal, including pre-emphasis and framing operations;

2) Carry out endpoint detection and cut processing to the voice signal preprocessed in step 1);

3) Check the effect of the voice signal after cutting in step 2), if the cutting is correct, save it in the specified folder;

4) When the recording of all speech signals in step 3) ends, a hidden Markov model of the speech signal is established, and the established model data is saved.

Preferably, in step 1), the pre-emphasis operation is to pass the signal through a high-pass filter to flatten the spectrum of the signal and keep it in the entire frequency band from low frequency to high frequency, and the spectrum can be calculated with the same signal-to-noise ratio.

Preferably, the endpoint detection method used in step 2) is a double-threshold detection method based on short-term energy and short-term zero-crossing rate, and the specific steps are:

1. Before starting the endpoint detection, first set two thresholds for the short-term energy and zero-crossing rate respectively, a low threshold with a small value, which is sensitive to signal changes and easy to be exceeded, and another high threshold with a numerical value Larger, the signal must reach the set strength before the threshold can be exceeded;

Two, the voice signal x (n) is carried out into frames, and each frame is recorded as n=1, 2, ..., N, n is a time series of discrete voice signals, N is a frame length, and i represents the number of frames;

3. Calculate the short-term energy of each frame of the speech signal, and obtain the short-time frame energy of the speech signal:

Where N is the frame length, i represents the number of frames, Represent the square of the value of the nth (1≤n≤N) sampling point of the speech signal of the i-th frame;

4. Calculate the zero-crossing rate of each frame of speech signal to obtain the short-term zero-crossing rate of the speech signal:

in:

Wherein sgn[s _i (n)] represents the value of the nth (1≤n≤N) sampling point of the speech signal of the i-th frame;

At this time, the entire endpoint detection is divided into four sections: silent section, transition section, speech section, and end section. In the silent section, if the short-term energy or zero-crossing rate exceeds the set low threshold, mark it as the starting point and enter the transition section Finally, if the two parameter values of short-term energy and zero-crossing rate fall back below the set low threshold, the current state will be restored to the silent section, and if the two parameters of short-term energy and zero-crossing rate in the transition section Any one of them exceeds the set high threshold, that is, it is considered to enter the speech segment. When in the speech segment, if the two parameter values of short-term energy and zero-crossing rate drop below the set low threshold, and the total timing length is less than If the minimum time threshold is used, it is considered as a section of noise, and the length of the speech signal is recalculated.

Preferably, in step 2), a threshold length is set according to the syllable of the speech signal and a variable silence1 that may be in the length of the speech segment signal is calculated. If the variable silence1 is less than the set threshold length, the loop is continued to recalculate the speech signal If the length is greater than the set threshold length, the previous speech signal will be discarded.

Preferably, the hardware of the place name speech modeling system oriented to goods sorting includes a high-performance noise-canceling earphone and a computer.

Preferably, the operation of the geographical name voice modeling system oriented to goods sorting can be completed on the operation interface by establishing an operation interface.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1, the present invention is aimed at the brief feature of place-name speech signal, regards each place-name speech signal as a recognition object, directly sets up the hidden Markov model with the whole of each place-name as an object, and calculation is simple.

2. The present invention establishes a graphical interface that is convenient for human-computer interaction, and is easy to operate. Through the system, you can directly see the voice effect map and the picture after the endpoint detection, and judge whether the detection is correct. When used in different regions, you can adjust the specific accent Modeling of a group of people to improve the accuracy of subsequent recognition.

Description of drawings

Fig. 1 is the principle diagram of the place name voice modeling system oriented to goods sorting in the present invention.

Fig. 2 is an effect diagram of an improved front-end point detection according to an embodiment of the present invention.

Fig. 3 is an effect diagram of endpoint detection after improvement according to the embodiment of the present invention.

Fig. 4 is the result figure of the double-threshold method endpoint detection of the speech signal of the embodiment of the present invention, wherein Fig. 4 (a) is the speech signal waveform diagram of the speech signal "Wuhan", and Fig. 4 (b) is the speech signal of the speech signal "Wuhan" Short-term energy waveform diagram, Figure 4(c) is the zero-crossing rate waveform diagram of the speech signal "Wuhan".

Fig. 5 is the hidden Markov chain model of the voice signal of place names oriented to goods sorting according to the present invention.

FIG. 6 is a schematic diagram of the present invention for establishing a hidden Markov model.

Fig. 7 is the interface of the place name voice modeling system oriented to cargo sorting of the present invention.

Fig. 8 is the user interface of the place name voice modeling system oriented to goods sorting according to the present invention.

detailed description

The present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

Example:

This embodiment provides an implementation method of a place-name phonetic modeling system oriented to cargo sorting, as shown in Figure 1 , which is a schematic diagram of a place-name phonetic modeling system oriented to cargo sorting in the present invention, and the method includes the following steps :

1) Preprocessing the voice signal, including pre-emphasis and framing operations;

In this step, the pre-emphasis operation is to pass the signal through a high-pass filter to flatten the frequency spectrum of the signal and keep it in the entire frequency band from low frequency to high frequency, so that the frequency spectrum can be calculated with the same signal-to-noise ratio. In this embodiment, the sampling frequency is 8KHz, the frame length is 256, and the frame shift is 128.

2) Carry out endpoint detection and cut processing to the voice signal preprocessed in step 1);

In this step, the endpoint detection method used is a double-threshold detection method based on short-term energy and short-term zero-crossing rate. The specific steps are:

1. Before starting the endpoint detection, first set two thresholds for the short-term energy and zero-crossing rate respectively, a low threshold with a small value, which is sensitive to signal changes and easy to be exceeded, and another high threshold with a numerical value Larger, the signal must reach the set strength before the threshold can be exceeded;

Two, the voice signal x (n) is carried out into frames, and each frame is recorded as n=1, 2, ..., N, n is a time series of discrete voice signals, N is a frame length, and i represents the number of frames;

3. Calculate the short-term energy of each frame of the speech signal, and obtain the short-time frame energy of the speech signal:

Where N is the frame length, i represents the number of frames, Represent the square of the value of the nth (1≤n≤N) sampling point of the speech signal of the i-th frame;

4. Calculate the zero-crossing rate of each frame of speech signal to obtain the short-term zero-crossing rate of the speech signal:

in:

Wherein sgn[s _i (n)] represents the value of the nth (1≤n≤N) sampling point of the speech signal of the i-th frame;

At this time, the entire endpoint detection is divided into four sections: silent section, transition section, speech section, and end section. In the silent section, if the short-term energy or zero-crossing rate exceeds the set low threshold, mark it as the starting point and enter the transition section Finally, if the two parameter values of short-term energy and zero-crossing rate fall back below the set low threshold, the current state will be restored to the silent section, and if the two parameters of short-term energy and zero-crossing rate in the transition section Any one of them exceeds the set high threshold, that is, it is considered to enter the speech segment. When in the speech segment, if the two parameter values of short-term energy and zero-crossing rate drop below the set low threshold, and the total timing length is less than If the minimum time threshold is used, it is considered as a section of noise, and the length of the speech signal is recalculated.

Since the signal of the present embodiment is a place name voice signal, each signal is composed of 2-4 syllables, when the first syllable is recognized, it may be because the second syllable is far away from the first syllable, and the first syllable The syllable length is too short, that is, the intermittent speech signal is directly filtered out as noise, so this embodiment sets a threshold length and a variable silence1 that may be in the length of the speech segment signal according to the syllable of the speech signal, if this variable silence1 is less than If the threshold length is set, the loop continues to recalculate the length of the voice signal, and if it is greater than the set threshold length, the previous voice signal is discarded. As shown in FIG. 2 , it is an effect diagram of the improved front-end point detection of this embodiment, and FIG. 3 is an effect diagram of the improved end point detection of this embodiment.

3) Check the effect of the voice signal after cutting in step 2), if the cutting is correct, save it in the specified folder;

As shown in Fig. 4, it is the speech signal double-threshold method endpoint detection result figure of the embodiment of the present invention, wherein Fig. 4 (a) is the speech signal waveform diagram of speech signal " Wuhan ", and Fig. 4 (b) is speech signal " Wuhan " The short-term energy waveform of the voice signal of ", Fig. 4(c) is the zero-crossing rate waveform of the voice signal of "Wuhan".

4) When the recording of all speech signals in step 3) ends, a hidden Markov model of the speech signal is established, and the established model data is saved.

Accompanying drawing 5 is the hidden Markov chain model of the voice signal of place name facing goods sorting of this embodiment, in this embodiment, a model includes 4 states, and state transition matrix is 4 * 4 matrix, and wherein element a _ij represents The probability of transitioning from state i to state j, states 1-4 contain four functions f ₁ , f ₂ , f ₃ , and f ₄ respectively. The four functions have the same structure, each of which consists of three 39-dimensional Gaussian probability densities Function composition, such as f ₁ , its structure is as follows:

f ₁ ＝k ₁₁ N[o,μ ₁₁ ,U ₁₁ ]+k ₁₂ N[o,μ ₁₂ ,U ₁₂ ]+k ₁₃ N[o,μ ₁₃ ,U ₁₃ ]

The above formula is the function f ₁ of state 1, the three coefficients represent the weight, and μ and U represent the mean and variance of the Gaussian probability density function. To sum up, a state in the model contains a 1×3 weight vector, a 3×39 mean matrix and a 3×39 variance matrix. Therefore, the data contained in a hidden Markov model are: a 4×4 state transition matrix, 4 1×3 weight vectors, 4 3×39 mean matrices and 4 3×39 variance matrices.

Accompanying drawing 6 sets up the schematic diagram of Hidden Markov Model for the present invention, at first input a class of voice signals, extract the MFCC (Mel Frequency Cepstral Coefficient) matrix of all signals, i.e. the characteristic matrix of signal, initialize the state transition of model Matrix and mean-variance matrix, use the Baum-Welch algorithm to re-evaluate the parameters, and then substitute all the characteristic matrices of this type of signal into the model, and use the model to calculate its output probability. If the growth rate of the output probability is less than the set If the threshold is set, the model has converged, the model training can be ended, and the data can be saved. If the growth rate is greater than the set threshold, the training will continue until the number of training times reaches the set number. When the model data is obtained, the model training ends, and the model data can be used for subsequent recognition procedures.

The specific operation is as follows: Accompanying drawing 7 shows the modeling interface of the place name speech recognition system. "System Operation Instructions" button, after clicking, the operation instruction interface will appear. First set up the "System Settings" dialog box, select the target folder, the gender of the operator, the total number of place name samples for this recording and the number of place name samples to be recorded. After the system settings are completed, in the "Sample Status" column, it will display Displays the name of the sample to be recorded and the number of existing samples. After the system setting is complete, you can operate the "System Operation" interface, click the "Record" button, and you can record with a duration of 2s. If it is correct, you can record the next sample. If it is wrong, click the "Delete" button to delete the sample recorded this time, and then perform the next recording. When all place name samples are recorded, you can click the "Build Voice Model" button to start modeling. After the modeling is completed, the system will automatically save the data to the selected folder, which is convenient for calling the voice recognition part. Accompanying drawing 8 is a schematic diagram during operation. In the "System Settings" column, after selecting the file path, the gender is set to "male", the total number of place name samples is 25, and the third place name sample will be recorded. In the "Sample Status" column It can be seen that the name of the third place name sample is "Beijing", and there is already 1 sample. In the "System Operation" column, samples can be recorded, and the status bar will display the sample information in time.

The above is only a preferred embodiment of the patent of the present invention, but the scope of protection of the patent of the present invention is not limited thereto. The equivalent replacement or change of the technical solution and its invention patent concept all belong to the protection scope of the invention patent.

Claims (6)

1. a kind of implementation method of the place name pronunciation modeling system towards goods sorting, it is characterised in that：Methods described include with Lower step：

1) pretreatment is carried out to voice signal, including preemphasis and framing operation；

2) to step 1) in pretreated voice signal carry out end-point detection and shear treatment；

3) check step 2) in voice signal after shearing effect, if shearing is correct, save it in the file specified In folder；

4) when step 3) in the admission of all voice signals terminate, set up the HMM of voice signal, and preserve The model data set up.

2. the implementation method of a kind of place name pronunciation modeling system towards goods sorting according to claim 1, its feature It is：Step 1) in, the preemphasis operation is, by high pass filter, to make the frequency spectrum of signal become flat by signal, is kept In the whole frequency band of low frequency to high frequency, frequency spectrum can be sought with same signal to noise ratio.

3. the implementation method of a kind of place name pronunciation modeling system towards goods sorting according to claim 1, its feature It is：Step 2) used in end-point detecting method be double-threshold comparison method based on short-time energy and short-time zero-crossing rate, specifically Step is：

First, before end-point detection is proceeded by, it is that short-time energy and zero-crossing rate set two thresholdings, a low door respectively first Limit, numerical value are less, and the change to signal is more sensitive, is easily exceeded, and another high threshold, numerical value are larger, and signal must reach The intensity of setting, the thresholding can be exceeded；

2nd, sub-frame processing is carried out to voice signal x (n), each frame is designated as n=1,2 ..., N, and n is discrete voice signal time sequence Row, N is frame length, and i represents frame number；

3rd, the short-time energy of each frame voice signal is calculated, the short time frame energy of voice signal is obtained：

$E_i=\underset{i=1}{\overset{N}{\Σ}}{s}_i^2\left(n\right)$

Wherein N is frame length, and i represents frame number,Represent n-th (1≤n≤N) individual sampled point of the i-th frame voice signal value it is flat Side；

4th, the zero-crossing rate of each frame voice signal is calculated, the short-time zero-crossing rate of voice signal is obtained：

$Z_i=\frac{1}{2}\underset{i=1}{\overset{N}{\Σ}}|\mathrm{sgn}\[s_i\left(n\right)\]-\mathrm{sgn}\[s_i(n-1)\]|$

Wherein：

Wherein sgn [s_i(n)] represent the i-th frame voice signal n-th (1≤n≤N) individual sampled point value；

Now whole end-point detection is divided into four sections：Quiet section, changeover portion, voice segments, ending segment, in quiet section when, if Short-time energy or zero-crossing rate exceed the low threshold of setting, are labeled as starting point, into after changeover portion, if short-time energy and zero passage Two parameter values of rate are all fallen back to below the low threshold of setting, and current state is returned to quiet section just, and if in changeover portion Any one in two parameter values of short-time energy and zero-crossing rate exceedes the high threshold of setting, that is, be considered as entering voice segments, be in During voice segments, if two parameter values of short-time energy and zero-crossing rate are all fallen below below the low threshold of setting, and total timing is long Degree is less than shortest time thresholding, then it is assumed that is one section of noise, recalculates the length of voice signal.

4. the implementation method of a kind of place name pronunciation modeling system towards goods sorting according to claim 3, its feature It is：Step 2) in, threshold value length is set according to the syllable of voice signal and a calculating is likely to be at voice segments letter The variable slience1 of number length, if this variable slience1 is continued cycling through, counted again less than threshold value length of setting The length of voice signal is calculated, if more than threshold value length of setting, voice signal above is given up.

5. the implementation method of a kind of place name pronunciation modeling system towards goods sorting according to claim 1, its feature It is：The hardware of the place name pronunciation modeling system towards goods sorting, including a high performance noise cancelling headphone and one Computer.

6. the implementation method of a kind of place name pronunciation modeling system towards goods sorting according to claim 1, its feature It is：The operation of the place name pronunciation modeling system towards goods sorting can be by setting up operation interface, in operation interface On complete.