CN1123862C - Speech recognition special-purpose chip based speaker-dependent speech recognition and speech playback method - Google Patents

Abstract

The present invention belongs to the field of voice technology. The present invention comprises the steps of voice recognition parameter extraction, training of specified person speech commands, recognition of the specified person speech commands and voice playback. The present invention has the characteristics of simple method, high recognition rate, good robustness, etc. Formed systems can be used for toy control, sound control dialing, intelligence domestic appliances and learning machines and can be used in control systems in a production link.

Description

Specific people's speech recognition, audio playback method based on the speech recognition special chip

Technical field the invention belongs to the voice technology field, relates in particular to 8 of employings or 16 monolithic MCU microcontrollers and realizes the specific people's audio recognition method of little vocabulary.

The specific people's speech recognition of background technology special chip, development is very fast abroad in recent years.More external voice technologies and semiconductor company all drop into a large amount of man power and materials and develop the speech recognition special chip, and oneself audio recognition method is carried out patent protection.The speech recognition performance of these special chips also has nothing in common with each other.Usually the process of speech recognition as shown in Figure 1, the voice signal of input is at first sampled through A/D, frequency spectrum shaping windowing pre-emphasis is handled, improve radio-frequency component, carry out real-time characteristic parameter extraction, the parameter of extraction is linear prediction cepstrum coefficient (LPCC) or Mel frequency marking cepstrum coefficient (MFCC), carry out end-point detection then, extract the efficient voice parameter, the lang sound recognition template of going forward side by side training or speech recognition template matches, and with best recognition result output playback.The hardware system of its special chip comprises 8 or 16 monolithic MCU microcontrollers and coupled automatic gain control (AGC), audio frequency preamplifier, low-pass filter, D/A (A/D), mould/number (D/A), audio-frequency power amplifier, voice operation demonstrator, random access memory (RAM), ROM (read-only memory) (ROM), width modulation (PWM) of execution speech recognition and phonetic synthesis, back method generally as shown in Figure 2.The speech recognition special chip RSC-164 series of products of U.S. Sensory company production at present are can buy one of best special chip of recognition performance in the world at present.These speech recognition special chips have been used for different mobile phones and wireless phone.Along with speech recognition technology improves, the speech recognition special chip will be widely used in various household electrical appliance and the control system, form information household electric industry, and this is one and develops rapidly and rising and high-tech industries that potentiality are very big.The mobile phone with specific people's speech recognition sound controlled dialing function of Philips company and Korea S Samsung release at present.The number of identification name is 10～20.Its recognition performance is unsatisfactory.

Summary of the invention the objective of the invention is for overcoming the weak point of prior art, a kind of specific people's speech recognition, audio playback method based on the speech recognition special chip proposed, can realize the specific people's speech recognition of high precision at cheap 8 monolithics or 16 MCU microcontrollers, it is low to have the method complexity, the high and good characteristics of robustness of accuracy of identification.Particularly Chinese digital voice and the recognition performance that easily mixes voice are reached even surpass current international most advanced level.

Specific people's speech recognition, audio playback method that the present invention proposes based on the speech recognition special chip, comprise the A/D sampling, frequency spectrum shaping windowing pre-emphasis is handled, characteristic parameter extraction, end-point detection, speech recognition template training and audio playback or speech recognition template matches are with best recognition result output playback, it is characterized in that, specifically may further comprise the steps:

A, speech recognition parameter extract:

(1) voice signal input back adopts A/D to sample, and becomes original digital speech, utilizes electric-level gain control, with the high precision of guaranteeing to sample;

(2) said original figure voice signal is carried out frequency spectrum shaping and divides the frame windowing process, to guarantee to divide the accurate stationarity of frame voice;

(3) minute feature of frame voice is carried out phonetic feature and extract, principal character is the cepstrum coefficient (LPCC) of linear prediction model (LPC) the computing voice feature according to voice, and storage is used for the dynamic segmentation and the template extraction step of back;

(4) use the zero-crossing rate and the short-time energy feature of voice signal to carry out end-point detection, remove the speech frame of no sound area, to guarantee the validity of each frame phonetic feature;

The training of B, specific people's voice command:

(1) phonetic feature that extracts is carried out dynamic segmentation and weighted mean, constitute template parameter, the parameter after the weighting is as new recognition template;

(2) this new template is carried out the identification feature analyzing and processing, guarantee new template and trained between the template that constitutes in the past to have the well property distinguished;

(3) to after handling, the voice that the property distinguished is bad, then prompting requires the speaker to re-enter new voice signal;

The identification of C, specific people's voice command:

(1) specific four steps of people's speech recognition process head are identical with said " speech recognition parameter extraction " process;

(2) this phonetic feature is compared with the recognition template of having stored, adopt Dynamic matching, extract the voice command that wherein mates most and as a result of export;

(3) in identifying, when recognition template matching error during, think that then recognition result is unreliable greater than certain thresholding, by prompting, require to re-enter voice.

G. audio playback:

The audio playback method adopts speech synthesis technique, said speech recognition parameter and phonetic synthesis model parameter are shared, with the speech recognition parameter while also as the phonetic synthesis model parameter, to reduce the expense of system as far as possible.

Electric-level gain control during said phonetic feature extracts can comprise: the input speech signal sampling precision is judged, if the input speech signal sampling precision is not high enough, by self-adaptive level control, adjusted the amplification quantity of voice, improve the speech sample precision; Said end-point detecting method is according to the end points thresholding of setting, and search for quiet section, determine voice, the top point.

Dynamic segmentation in the training of said voice command and weighted average method specifically can may further comprise the steps:

(1) at first according to the variation of parameter between speech characteristic parameter computing voice different frame, surpasses a certain setting threshold, determine that this frame is an important separation in the phonetic feature when changing;

(2) number to its separation of different phonetic signal can be different; Phonetic feature between the different separations is weighted on average, improves the proportion of important phonetic feature in model of cognition.

Said identification feature analytical approach specifically can comprise: relatively new template is with the Dynamic matching distance threshold between the old template, when thresholding during greater than a certain definite statistical value, new template stores as recognition template, otherwise thinks that this template is invalid, requires to re-enter voice signal.

The method that identification parameter in the said audio playback and voice coding channel model parameter are shared specifically can may further comprise the steps:

(1) the speech recognition modeling parameter adopts identical parameter with the voice coding channel parameters, does not therefore need to increase the memory space of channel model parameter in speech.

(2) excitation parameters of channel model adopts improved LPC vocoder method, and excitation parameters is a pitch period, clear/turbid/transition sound determination information.

The present invention has following characteristics:

(1) the present invention is based on the specific people's audio recognition method of the medium and small vocabulary of speech recognition special chip.These methods have characteristics such as complicacy is low, accuracy of identification is high, robustness is good.Be particularly suitable for very limited 8 8-digit microcontrollers of arithmetic capability.

(2) adopt the shared way of identification parameter and coding parameter, thereby significantly reduced requirement, guarantee to have very high coding quality simultaneously system resource.

(3) compared with the prior art has the better recognition performance to obscuring the vocabulary book chip easily.

(4) owing to adopt 8 or 16 MCU cores, 10 bit linear A/D, D/A, so outstanding feature such as this chip has that volume is little, in light weight, power consumptive province, cost are low.In fields such as communication, Industry Control, intelligent home electrical appliance, intelligent toy, automotive electronics great using value is arranged.

(5) the present invention is 30 to 8 MCU recognition commands, is 60 to the order of 16 bit DSP chip identification.Phonetic recognization rate to 8 chips is 95%, is 99% to the phonetic recognization rate of 16 bit DSP chips.

Description of drawings

Fig. 1 is the process schematic block diagram of common speech recognition.

Fig. 2 is that the hardware system of general voice special chip is formed synoptic diagram.

Fig. 3 totally constitutes synoptic diagram for the method for the embodiment of the invention.

The end-point detecting method block diagram of Fig. 4 present embodiment as shown.

Fig. 5 is the voice training process overall flow block diagram of present embodiment.

Fig. 6 is the speech recognition process overall flow block diagram of present embodiment.

The specific people's speech recognition based on the speech recognition special chip that embodiment the present invention proposes, the embodiment of audio playback method are described in detail as follows in conjunction with each figure:

The overall formation of present embodiment as shown in Figure 3, whole process can be divided into the pre-emphasis of (1) A/D sampling and sampling back voice, improves the energy of high-frequency signal, windowing divides frame to handle; (2) extraction of speech characteristic parameter (comprising end-point detection parameter, model of cognition parameter); (3) end-point detection is determined effective speech parameter; (4) effective speech characteristic parameter is carried out dynamic segmentation; (5) specific people's voice are carried out the training of template and the extraction of voice playback synthetic parameters; (6) speech recognition is carried out template relatively by method for mode matching; And with voice identification result by the audio playback technology export.The specification specified of each step is as follows.

1, speech recognition parameter feature extraction:

(1) voice signal at first carries out low-pass filter, samples by 10-bit linear A/D then, becomes original digital speech, and adopting the purpose of 10 A/D is in order to reduce the cost of chip.Because the precision of A/D is low, therefore to control and the energy and the overload situations of input signal are judged gain-controlled amplifier on the method, so that guarantee to have made full use of the dynamic range of 10 A/D, obtain high as far as possible sampling precision.

(2) the original figure voice signal is carried out frequency spectrum shaping and divides the frame windowing process, the accurate stationarity that guarantees to divide the frame voice.Preemphasis filter is taken as 1-0.95z ^-1, zero-crossing rate lifts level and is taken as 4 in calculating.

(3) minute feature of frame voice is carried out phonetic feature and extract, phonetic feature comprises LPCC cepstrum coefficient, energy, zero-crossing rate etc., and storage is used for the back dynamic segmentation.The calculating of a wherein very important step correlation function value need be finished in real time, owing to based on 8 single-chip microcomputer 8 no sign multiplication is only arranged, the process of therefore calculating correlation function value is as follows: $a\left(n\right)=s\left(n\right)+128$ $R\left(i\right)=\underset{n}{\mathrm{\Σ}}s\left(n\right)\×s(n+i)=\underset{n}{\mathrm{\Σ}}(a\left(n\right)-128)\×(a(n+i)-128)$ $=\underset{n}{\mathrm{\Σ}}a\left(n\right)\×a(n+i)-128\×\underset{n}{\mathrm{\Σ}}(a\left(n\right)+a(n+i))+\underset{n}{\mathrm{\Σ}}128\×128$

In the following formula, s (n) is converted into unsigned number a (n) for 8 signed numbers are arranged.Obviously product is with three words

Joint is preserved and can not be overflowed (frame length is not more than 256).

2, end-point detection:

(1) guarantees the validity of each frame phonetic feature, eliminate irrelevant noise, must carry out the end-point detection and the judgement of voice.End-point detecting method of the present invention was divided into for two steps, at first end points is carried out preliminary ruling, after energy is greater than a certain determined value, be defined as preliminary starting point according to speech signal energy, continue to seek the bigger unvoiced frame of speech signal energy backward from this starting point then, carry out the voiced segments location.Be in the main true if unvoiced frame exists this end points of explanation to judge, begin to search for forward, backward the start frame of quiet frame as voice from unvoiced frame.Result's output with search.The end-point detection block diagram as shown in Figure 4.Basic skills is described below: ZERO_RATE_TH is a threshold value of zero-crossing rate, and ACTIVE_LEVEL, INACTIVE_LEVEL and ON_LEVEL are the threshold values of energy.

(2) initial value of system is decided to be silent state.Under silent state, when zero-crossing rate surpasses threshold value ZERO_RATE_TH or energy and surpasses threshold value A CTIVE_LEVEL, change state of activation over to, if energy surpasses threshold value ON_LEVEL, then directly change sonance over to.Remember that this frame is the forward terminal of voice.

(3) under state of activation,, then change sonance over to if energy surpasses threshold value ON_LEVEL; If continuous some frames (being set by constant C ONST_DURATION) energy all surpasses only threshold value ON_LEVEL, change no voice and spirit over to.

(4),, then change unactivated state over to if energy is lower than threshold value INACTIVE_LEVEL at sonance.This frame of mark is the aft terminal of voice.

(5) in unactivated state, if continuous some frames (being set by constant C ONST_DURATION) energy all surpasses only threshold value INACTIVE_LEVEL, then voice finish; Otherwise change sonance over to.

The actual value of parameter is as follows: ZERO_RATE_TH is taken as 0.4, and ACTIVE_LEVEL is more according to the background noise setting, and INACTIVE_LEVEL is taken as 4 times of ACTIVE_LEVEL, and ON_LEVEL is taken as 8 times of ACTIVE_LEVEL, and CONST_DURATION is made as 20 frames.

3, phonetic feature dynamic segmentation, weighted mean:

(1) the input phonetic feature is carried out dynamic segmentation and weighted mean, improve the proportion of voiceless consonant characteristic parameter in identification, extract most important template parameter in the phonetic feature.The phonetic feature segmentation is one of core of this system voice recognition methods.

(1) the normalization Euclidean distance of calculating the speech characteristic parameter between different frame is adopted in dynamic segmentation.Surpass certain thresholding when changing, assert that this point is the important separation of phonetic feature.Phonetic feature in the different sections is weighted on average, and they are preserved as new speech characteristic parameter, and remove previous phonetic feature.By model parameter is reduced widely, not only save storage space, and reduced the complexity of computing and improved system's arithmetic speed.

4, the training of specific people's speech recognition template:

(1) before carrying out specific people's speech recognition, at first to train the system identification template.On the basis that three step speech characteristic parameters extract, the training study process will carry out twice, can improve the robustness of model of cognition parameter so in front.The template of setting up in the characteristic parameter of extraction in the training for the second time and the training is for the first time carried out dynamic programming, find out corresponding segment information, be weighted average then, model of cognition parameter as final candidate, carry out the identification feature analyzing and processing of system at last, have the well property distinguished between the recognition template of guaranteeing new template and training in the past, likelihood ratio should be greater than 1.6 between the template.Can not cause damage like this to the system identification performance.Different templates is carried out necessary adjustment, increase the separability of different templates.

(2) but to the phonetic feature that the distinguishing characteristic of adjusting between the rear pattern plate does not still meet the demands, then according to circumstances, require the speaker to retell this same voice by voice suggestion, increase frequency of training, or the suggestion user imports new different phonetic.Training managing by these two steps can make system have good identification feature, keeps very high discrimination.

5, specific people's speech recognition:

(1) the identifying feature extraction is identical with the front feature extracting method.

(2) phonetic feature is compared with the template of having stored, its computation process adopts the nonlinear dynamic programming method for mode matching of speech recognition, searching and certain each specific template are near the output of voice command as voice identification result, for improving the system identification reliability when finally exporting as a result, also to carry out confidence level and the calculating of refusing to know model.

(3) the credible calculating of estimating and refusing to know: with first likelihood ratio of selecting identification probability and first three to select the average probability of recognition result to constitute, and first the likelihood ratio of selecting identification probability and second to select probability to constitute be combined into and comprehensive crediblely estimate, if this likelihood ratio is less than 3.0, then think credible estimate low, recognition result is uncertain voice or noise, and it is refused to know, voice are re-entered in prompting; Estimate highly for credible, then export recognition result.Can eliminate the interference of neighbourhood noise by refusing to know processing to recognition system.

6, audio playback is handled:

(1) audio playback is handled the encoding and decoding speech method that adopts usually.The speech recognition modeling parameter adopts identical parameter with the voice coding channel parameters, does not therefore need to increase the memory space to the voice coding model parameter in speech.The encoding and decoding speech model is improved LPC vocoder.

(3) excitation parameters of channel model adopts improved LPC vocoder method, and excitation parameters is a pitch period, clear/turbid/transition sound determination information.In order to improve the voice coding quality, should be in decode procedure with front and back frame speech channel parameter, excitation parameters carries out linear interpolation, improves seamlessly transitting between the voice between different frame.

The voice training process overall flow of present embodiment as shown in Figure 5, at first the recognition system prompting input voice first time then carry out end-point detection and feature extraction, and carry out dynamic segmentation, constitute initial recognition template, system prompt is imported voice for the second time, then carries out end-point detection and feature extraction, utilizes dynamic programming method, carry out Dynamic matching with the initial identification template, find out segment information, carry out arithmetic mean then, constitute new recognition template.Identifiability between the template of judging new recognition template and training in the past is good to getting off as template stores to distinctive.Voice are re-entered in the requirement that distinctive is bad.

The speech recognition process overall flow of present embodiment as shown in Figure 6, at first import voice, then carry out end-point detection and feature extraction, each template of storing in this phonetic feature and the system is carried out dynamic programming, pattern match, and will mate best three recognition results as output, and carry out the credible calculating of estimating, if crediblely estimate, get and crediblely estimate the highest template and export as recognition result greater than definite thresholding.If crediblely estimate less than definite thresholding, system refuses to know.

Present embodiment has been developed a kind of speech recognition special chip based on the specific people's audio recognition method of medium and small vocabulary of above-mentioned sound identification special chip and has been comprised: audio frequency prime amplifier, automatic gain control (AGC), D/A (A/D) converter, mould/number (D/A) converter, MCU nuclear (8051), pulse width modulator (PWM), random access memory (RAM), ROM (read-only memory) (ROM), flash memory (FLASH).Store phoneme synthesizing method, voice coding method, speech recognition training method and audio recognition method among the ROM.The template and the voice playback of speech recognition are stored among the FLASH.

Claims (5)

1, a kind of specific people's speech recognition, audio playback method based on the speech recognition special chip, comprise the A/D sampling, frequency spectrum shaping windowing pre-emphasis is handled, characteristic parameter extraction, end-point detection, speech recognition template training and audio playback or speech recognition template matches are with best recognition result output playback, it is characterized in that, specifically may further comprise the steps:

A, speech recognition parameter extract:

(1) voice signal input back adopts A/D to sample, and becomes original digital speech, utilizes electric-level gain control, with the high precision of guaranteeing to sample;

(2) said original figure voice signal is carried out frequency spectrum shaping and divides the frame windowing process, to guarantee to divide the accurate stationarity of frame voice;

(3) minute feature of frame voice is carried out phonetic feature and extract, principal character is the cepstrum coefficient (LPCC) of linear prediction model (LPC) the computing voice feature according to voice, and storage is used for the dynamic segmentation and the template extraction step of back;

(4) use the zero-crossing rate and the short-time energy feature of voice signal to carry out end-point detection, remove the speech frame of no sound area, to guarantee the validity of each frame phonetic feature;

The training of B, specific people's voice command:

(1) phonetic feature that extracts is carried out dynamic segmentation and weighted mean, constitute template parameter, the parameter after the weighting is as new recognition template;

(2) this new template is carried out the identification feature analyzing and processing, guarantee new template and trained between the template that constitutes in the past to have the well property distinguished;

(3) to after handling, the voice that the property distinguished is bad, then prompting requires the speaker to re-enter new voice signal;

The identification of C, specific people's voice command:

(1) specific four steps of people's speech recognition process head are identical with said " speech recognition parameter extraction " process;

(2) this phonetic feature is compared with the recognition template of having stored, adopt Dynamic matching, extract the voice command that wherein mates most and as a result of export;

(3) in identifying, when recognition template matching error during, think that then recognition result is unreliable greater than certain thresholding, by prompting, require to re-enter voice.

D. audio playback:

The audio playback method adopts speech synthesis technique, said speech recognition parameter and phonetic synthesis model parameter are shared, with the speech recognition parameter while also as the phonetic synthesis model parameter, to reduce the expense of system as far as possible.

2, specific people's speech recognition as claimed in claim 1, audio playback method, it is characterized in that, electric-level gain control during said phonetic feature extracts comprises: the input speech signal sampling precision is judged, if the input speech signal sampling precision is not high enough, control by self-adaptive level, adjust the amplification quantity of voice, improve the speech sample precision; Said end-point detecting method is according to the end points thresholding of setting, and search for quiet section, determine voice, the top point.

3, specific people's speech recognition as claimed in claim 1, audio playback method is characterized in that dynamic segmentation in the training of said voice command and weighted average method specifically may further comprise the steps:

(1) at first according to the variation of parameter between speech characteristic parameter computing voice different frame, surpasses a certain setting threshold, determine that this frame is an important separation in the phonetic feature when changing;

(2) to the number difference of its separation of different phonetic signal; Phonetic feature between the different separations is weighted on average, improves the proportion of important phonetic feature in model of cognition.

4, specific people's speech recognition as claimed in claim 1, audio playback method, it is characterized in that, said identification feature analytical approach specifically comprises: relatively new template is with the Dynamic matching distance threshold between the old template, when thresholding during greater than a certain definite statistical value, new template stores as recognition template, otherwise think that this template is invalid, require to re-enter voice signal.

5, specific people's speech recognition as claimed in claim 1, audio playback method is characterized in that, the method that identification parameter in the said audio playback and voice coding channel model parameter are shared specifically may further comprise the steps:

(1) the speech recognition modeling parameter adopts identical parameter with the voice coding channel parameters, does not therefore need to increase the memory space of channel model parameter in speech.

(2) excitation parameters of channel model adopts improved LPC vocoder method, and excitation parameters is a pitch period, clear/turbid/transition sound determination information.

Images