Biometric technology is the use of human biometric identity authentication technology that is currently recognized as the most convenient and secure identification technology.
Because each person's biometric with others different unique and in a given period of stability, it is difficult to counterfeit and fraudulent identification, so take advantage of the bodies and technology for identity authentication, secure, accurate, and reliable.In the field of bio-recognition, voice recognition, also known as speaker recognition, with its unique convenience, efficiency and accuracy, and other advantages are impressive, and I become the people's daily lives and work in an important and common security authentication mode.
Voice recognition is a speaker speech waveform reflected speaker physiological and behavioral characteristics of the voice parameters, automatic identification identity technologies described.Voice recognition technology can be divided into two categories, namely, speaker identification/speaker verification.
The former is used to determine a pronunciation is the number of people that is a multiple-selection of a problem; the latter used to confirm a voice is number one in which one was more a problem; the latter used to confirm whether a voice is designated one person said, the issue of discrimination is one-to-one. On the other hand, there is also the voice recognition and text and text-independent two, according to the specific tasks and application, the application scope. And the text of voice recognition system requires that the user content in accordance with the provisions the pronunciation, everyone's VoicePrint model was precisely to establish individually, but also must be identified according to the rules of pronunciation, so you can achieve better recognition; and text-independent identification system does not provide speaker sound content, model relatively difficult, but the user easy to use, a wide range of applications.This article describes the voice electronic door lock is a 16-bit MCU Sunplus SPCE061A on implementation and text for speaker verification system.
This system is mainly composed of speaker recognition module, door lock control motor and door locks. In training, the speaker's voice enters through the microphone speaker speech signal acquisition front-end circuit, speech signal processing circuit on the acquisition of voice signal characteristics and speech processing, extracting the speaker's personality features and storage that form the speaker feature parameters of the database. In recognition, the question to identify voice and speaker feature parameters to match the database, via the output circuit control door motor, and ultimately on the door of the control.1 algorithm theory
Speaker recognition algorithm principle diagram shown in Figure 1.
1.1 pretreatment
(1) noise
On the microphone input analog voice signal to quantization and sampling, and access to digitized voice signals; and noisy voice signal denoising, get clean voice signal and preemphasis technology to filter out low-frequency interference, in particular, 50 Hz or 60 Hz frequency interference, voice signal of high frequency components, but it can also serve to eliminate DC drift, the random noise suppression and enhancing clarity on the role of some energy.
(2) endpoint detection
This system uses the voice signals of short-term energy and short-time zero rate for endpoint detection.
Voice signal sampling frequency for each frame data 8kHz, 20ms, for a total of 160 sampling points. Every 20ms calculated once short-term energy and short-time zero rate. Through the speech signal of short-term energy and short-time zero rate detection can eliminate fall silent frame, white noise and sound between frames, and finally retaining for pitch, LPCC, feature very useful voiced sound signal.1.2 feature extraction
After the voice signal pre-processing, followed by extraction of the characteristic parameters.
Feature extraction task is to extract the voice signal characterization of basic characteristics.1.2.1 feature parameter selection
Features must be able to effectively distinguish different speaker, and on the same speaker changes remained relatively stable, at the request of a simple calculation of parameters, it is best to have efficient algorithm to ensure recognition of real-time.
Speaker features generally fall into the following categories:
(1) based on the vocal organs such as the glottis, channel and nasal structure and extraction of physiological parameters.
As spectrum envelope, pitch, formant, etc. One pitch well depicting the sound characteristics of the speaker, to a large extent reflects a person's personality.(2) model based on channel characteristics, linear predictive analysis of the parameters.
Including linear prediction coefficients (LPC) and linear prediction exported to various parameters, such as linear prediction cepstrum coefficient (LPCC), part of the correlation coefficient, the reflection coefficient, log area ratio, LSP line spectrum on, linear prediction residuals. According to the results of the work of their predecessors and actual test comparison, LPCC parameters not only better feedback channel's formant characteristics, with better recognition results, and can be used relatively simple operation and a faster speed.(3) based on the ear hearing mechanism, reflect the auditory characteristics, analog voice frequency ear on the perception of the characteristic parameters.
Such as the United States-cepstrum coefficient (MFCC), etc. MFCC parameters based on linear prediction of Cepstral analysis compared to prominent advantage is not dependent on the entire extreme voice generation model assumed that and Guangxi-independent speaker recognition system MFCC parameter to LPCC parameter to better improve the system of recognition performance.In addition, it was also through the different characteristic parameters of combinations to increase the actual system performance.
When the combination of parametric correlation is not large, it will have better results, because they each reflect the different characteristics of speech signal.Simulation of the computer platform, through the various parameters of the actual comparison, MFCC parameters than using LPCC parameters have better recognition results.
But do the SPCE061A platform for real-time processing, and LPCC system, MFCC coefficient calculation has two drawbacks: first, the calculation for a long time; the second is fineDegree of difficulty. As a result of the calculation needed MFCC system FFT transform and logarithmic operations affecting the calculation of the dynamic range; to ensure the system to recognize only the real-time, sacrifice precision parameters. And there are parameters calculation LPCC recurrence formula, speed and precision can guarantee, recognition also meet real needs.This system incorporates the pitch and LPC cepstrum coefficient (LPCC) as speaker recognition feature parameters.
1.2.2 LPCC parameter extraction
Based on linear prediction analysis of Cepstral parameter LPCC can simple recursion formula consists of linear prediction coefficients evaluated.
Recurrence formula is as follows:Where p is the number of order LPC model is the model number of poles.
(1) LPC model order p of
In order to make the model assumes that better meet the voice generation model, you should make the order of LPC model p and formant numbers match, followed by considering the glottis pulse shape and face the effects of radiation.
Usually a pair of Poles corresponds to a formant, 10kHz sampling of speech signal typically have five formant, p = 10, 8kHz sampling of speech signal desirable p = 8. In addition to compensate for the nasal in zero, as well as other factors, usually caused by deviations in the above order on the basis of the two poles, which are, respectively, p = 12 and p10. Experiments show that select order LPC analysis p = 12, for the vast majority of voice signal channel model approximation can be sufficient approximation. P value selected too large can slightly improved approximation effect, but also has some negative effects, on the one hand, it is necessary to intensify the calculated measures, on the other hand has the potential to add some unnecessary details.(2) linear prediction coefficients of
[1] [2] [3]
No comments:
Post a Comment