CN1342968A - High-accuracy high-resolution base frequency extracting method for speech recognization - Google Patents

Abstract

The high-precision and high-resolution fundamental frequency extraction method for speech recognition is a fundamental frequency extraction method that combines frequency domain, time domain analysis and dynamic programming (DP). The main features are: firstly perform FFT transformation on the voice signal in the frequency domain, then perform harmonic analysis on it, and then select several candidates for the fundamental frequency through peak detection, and use the autocorrelation coefficient for these candidate fundamental frequencies in the time domain Then use the dynamic programming algorithm to look back at an indefinite length to determine an optimal fundamental frequency contour based on the analysis results of the frequency domain and time domain and the variation of the fundamental frequency. In order to ensure the resolution of fundamental frequency extraction, methods such as downsampling rate and interpolation are adopted.

Description

The high-accuracy high-resolution fundamental frequency extracting method that is used for speech recognition

The present invention has extracted a kind of fundamental frequency extracting method that frequency domain, time-domain analysis and dynamic programming (DP) are combined, belong to signal Processing and field of speech recognition, its essential characteristic is the fundamental frequency analysis in conjunction with time-domain and frequency-domain, extract possible several fundamental frequency candidates, and the objective function that adopts a kind of coefficient of autocorrelation, spectrum analysis result and fundamental frequency variable quantity to combine is recalled the search that dynamic programming algorithm is carried out correct fundamental frequency with random length.

At present, under the continuous speech condition, extracting reliable fundamental frequency change curve is the basis of Chinese language tone identification and modeling.But the difficulty that fundamental frequency extracts mainly is, is subjected to the influence of factors such as sound channel resonance peak and noise, and frequency multiplication and the half frequency value at correct fundamental frequency has than the more obvious feature of fundamental frequency (bigger as corresponding peaks) sometimes.General independent time domain or the frequency domain of utilizing of traditional method carries out the fundamental frequency feature extraction, and only extracts the basis of a single fundamental frequency value as subsequent treatment such as median smoothing at all time points.So only with a kind of method of frequency domain or time domain, only extract fundamental frequency and some mistakes usually occur from an extreme standard (as only getting a certain largest peaks), cause the pitch contour line that obtains that obvious discontinuous variation is arranged, this does not obviously square with the fact.As: the test findings that J.Sundberg was published on the J.of Phonetics in 1979 shows that the rate of change of fundamental frequency is in the scope of about 1%/ms.Obvious discontinuous variation for fundamental frequency, can be with certain the remedying in addition of traditional DP method, but must handling one whole section voice, traditional DP algorithm could export the pitch contour line, this delay is very difficult received in real-time speech recognition system, all are in the application of some other DP algorithm, and the method that all adopts fixed length to recall addresses this problem.

The subject matter that fundamental frequency extracts is:

1. on frequency domain, the frequency spectrum that voice signal obtains through FFT (fast fourier transform) is at F ₀/ 2, F ₀, 2F ₀, 3F ₀The place often has bigger peak to occur, as shown in Figure 2, the spectrum that a unvoiced frame obtains after the FFT computing, it is at F ₀, 2F ₀, 3F ₀There is tangible peak at the place, but the peak value at fundamental frequency F0 place does not obviously have 2F ₀The place is high.If any algorithm just be based on this phenomenon, the frequency of maximum peak correspondence as fundamental frequency; Perhaps someways extract a plurality of peaks, and calculate between these peak values range averaging with the estimation fundamental frequency.But some peak value is not fairly obvious sometimes, even the phenomenon of " fundamental tone disappearance " (missing fundamental) can occur, and this has caused very big difficulty for traditional method of doing fundamental detection with spectrum analysis.

2. on time domain, a kind of representational fundamental frequency extracting method is the short-time autocorrelation function method.The computing formula of short-time autocorrelation function is: $R\left[m\right]=\frac{1}{N-m}\underset{i=0}{\overset{N-m-1}{\mathrm{\Σ}}}\mathrm{xixi}+m$ X wherein _iRepresent i sampling point value in the frame, N represents the total sampled point number of every frame, R[m] autocorrelation value when indication cycle is m sampled point.According to the character of autocorrelation function, the autocorrelation function of cyclical signal also is periodic, and identical with the cycle of voice signal, in general, if T ₀Represent pitch period (T ₀=1/F ₀), except R[0], T ₀The peak value maximum at place.Fig. 4 has provided the autocorrelator trace of a unvoiced frame, can see, at T ₀, 2T ₀, 3T ₀There is bigger peak value at the place, but T ₀The peak height and the 2T at place ₀, 3T ₀Very approaching.So the maximum peak that detects autocorrelation function can be used as a kind of method of estimating fundamental frequency, but similar with top situation, at T ₀/ 2,2T ₀, 3T ₀Near bigger peak value is usually arranged, and in these peaks maximum peak appears sometimes.This brings certain limitation for this algorithm.

In addition, resolution and calculated amount have caused certain difficulty also for the extract real-time of fundamental frequency.

For example in the method that adopts the FFT spectrum analysis, the resolution that fundamental frequency extracts is directly proportional with counting of FFT calculating, and its calculated amount is to be exponential increase thereupon, and this just causes and can not calculate with the FFT that counts too much; (the fundamental frequency scope that refers to human speech) calculates auto-correlation with a large amount of multiplying of search higher value needs in all effective frequency domain scope, is very consuming time, needing not too to be suitable in the speech recognition system of processing in real time; The candidate value number of DP algorithm also influences the speed of decision of DP, if with all fundamental frequencies in effective frequency domain scope all as the candidate, also certainly will increase the weight of the burden of making a strategic decision greatly.

The objective of the invention is to: propose a kind of new fundamental frequency extracting method.The basic thought of this method should be: earlier voice signal is carried out the FFT conversion on frequency domain, again it is carried out frequency analysis, detect by peak value then, select the candidate of several fundamental frequencies, on time domain, these candidate's fundamental frequencies are evaluated and tested with coefficient of autocorrelation, determined the pitch contour line of an optimum again with dynamic programming algorithm.

Technical essential of the present invention is: a kind of speech recognition fundamental frequency analysis of high-accuracy high-resolution fundamental frequency extracting method in conjunction with time-domain and frequency-domain, extract possible several fundamental frequency candidates, and adopt a kind of coefficient of autocorrelation, the objective function that spectrum analysis result and fundamental frequency variable quantity combine is searched for dynamic programming algorithm, before carrying out auto-correlation calculating, the spectrum analysis result who uses according to speech recognition, utilize the FFT harmonic analysis method, fundamental frequency scope to voice is carried out preliminary estimation, reduce the scope that auto-correlation is calculated, and fundamental frequency extraction precision is not exerted an influence; The feature of harmonic analysis method comprises down-sampled rate, pre-emphasis and windowing, FFT calculating and carries out steps such as secondary Spline interpolation in order to improve resolution.1.FFT the frequency spectrum of trying to achieve adopts harmonic analytic method, at first for each fundamental frequency of effective fundamental frequency scope calculate adding up of several harmonic waves and.If use X _(i)The voice signal of representing a frame is represented its FFT conversion with X (f), the power spectrum that S (f) expression obtains, that is:

S (f)=| X (f) | ²(2) result that adds up of harmonic wave is so: $H\left(f\right)=\underset{n=1}{\overset{\mathrm{HN}}{\mathrm{\Σ}}}{h}_{n}S\left(\mathrm{nf}\right)---\left(3\right)$ HN represents the number of harmonic wave, h _nBe the power of n harmonic wave.Power spectrum is not carried out logarithm Log computing in this method, linear frequency domain is not transformed into the logarithm frequency domain yet, promptly ask Log (F ₀), this has reduced calculated amount to a certain extent.And then from the H as a result (f) of frequency analysis, select the peak of top's maximum by linear search, (establish P ₁, P ₂... P _PNRepresent this several peak values), with the candidate value of correspondent frequency as fundamental frequency.Calculate the relative maximum peak P of each peak value simultaneously _MaxSize as the grading parameters H of back DP algorithm _Per, that is: $H_{\mathrm{per}}\left(i\right)=\mathrm{Pi}/P_{\max },i=1..\mathrm{PN}---\left(4\right)$ The harmonic wave that example such as Fig. 3 that frequency analysis fundamental frequency peak value detects, Fig. 3 have provided the unvoiced frame figure that adds up is through Harmonics Calculation, at F ₀There is a tangible maximum peak at the place, together with the several peaks (representing with circle) that also have other in addition, all as the candidate of fundamental frequency.2. on time domain, the fundamental frequency candidate that frequency domain is obtained calculates autocorrelation value { R[i] }, its result R[0] carried out normalization, i.e. coefficient of autocorrelation: R _Per(i)=R[i]/R[0].As mentioned above, the peakedness ratio of the coefficient of autocorrelation of voiced sound (voiced) frame is bigger, with 1 more approaching, as shown in Figure 4, and do not have the autocorrelative peak value of non-voiced sound (unvoiced) frame of fundamental frequency all very as shown in Figure 5 low, so coefficient of autocorrelation can be used as the foundation that pure and impure sound is differentiated, and can filter out some tangible non-unvoiced frames.Utilize coefficient of autocorrelation, the limited fundamental frequency candidate that frequency domain is obtained gives a mark and screens, both reduced the calculated amount of autocorrelative calculated amount and follow-up DP algorithm, also made simultaneously when selecting fundamental frequency the comprehensively analysis result of frequency domain and time domain with DP algorithm, more reasonable.3. dynamic programming algorithm is a kind of widely used optimization algorithm.It turns to the subprocess of a plurality of single phases decision-making to multistage decision process, makes computational short cut.Each decision-making is given a mark respectively according to scoring function to all possible paths, according to the size of score the path is reduced, and exports the path of score optimum (maximum or minimum) at last.The each fundamental frequency candidate who handles a frame of this DP algorithm carries out continuation to original fundamental frequency path, and to each paths marking, it is the highest to keep some scores then.The score formula is:

Score (i)=max{Score (i-1)-D (i, j) }+aR _Per(i)+bH _Per(i) (5) wherein Score (i) represent the score of this fundamental frequency path at the i frame, a, b are coefficient, D (i, j) the fundamental frequency p of expression i frame _iJ fundamental frequency candidate p with the i-1 frame _jDistance:

D (i, j)=2 ^*| p _i-p _j|/(p _i+ p _j) the ∑ R that adopts of (6) this method _Per(i) scoring function has been considered three ∑ H _Per(i) factor of aspect: continuity, coefficient of autocorrelation accumulated value maximum, the harmonic wave that fundamental frequency the changes peak value maximum that adds up.These three aspects all are the important evidence that fundamental frequency is differentiated.Latter two factors was mentioned in front.Wherein the continuity of fundamental frequency variation is that the fundamental frequency track of considering human pronunciation can not produce the principle of suddenling change, to possible fundamental frequency path marking.This algorithm does not have the characteristics of obvious fundamental frequency according to non-voiced sound, has adopted the random length retrogressive method.

Concrete grammar is as follows: first frame with continuous unvoiced frame of this DP algorithm begins, and does end with the last frame of continuous unvoiced frame, carries out the part of dynamic random length and recalls.Because the fundamental frequency that proposes at non-unvoiced frame is nonsensical, the continuity of fundamental frequency also can't embody after entering unvoiced frame.If in the overall situation, carry out dynamic programming, then certainly will add too much interference mark, make pitch contour line generation random offset, have a strong impact on the effect of DP algorithm sometimes.The basis for estimation zero-crossing rate of continuous unvoiced frame detects and coefficient of autocorrelation in this method, chosen a zero-crossing rate fault value and a coefficient of autocorrelation fault value that better discrimination is arranged according to statistics, parameter is judged as non-voiced sound less than the frame of this fault value, does not participate in DP algorithm and asks the pitch contour line.Reducing major part does not like this have the non-unvoiced frame of fundamental frequency, greatly reduces their influences to extracting fundamental frequency with DP algorithm, has reduced mistake; Both guarantee simultaneously the continuity of pitch contour line, also significantly reduced the delay of DP algorithm.The non-unvoiced frame of minority of being included can not have big influence to DP algorithm; The unvoiced frame that minority is not included can well be compensated in line.4. other measure.Zero-crossing rate detects.As mentioned above, for differentiation rate that improves voiced sound and non-voiced sound and the speed that improves the fundamental frequency extraction, voice signal has at first been carried out the zero-crossing rate detection.Preestablish a bigger fault value, zero-crossing rate surpasses the non-voiced sound that is judged as of this value, does not carry out fundamental frequency and extracts.Otherwise extract fundamental frequency with FFT/DP algorithm cited below.The bigger reason that the fault value is established is to guarantee that voiced sound is not filtered.In order to reduce the influence of waveform skew, the reference value of zero-crossing rate is not zero, but the average of frame data.Down-sampled rate and interpolation.Reduce fundamental frequency extraction precision for improving, carried out down-sampled rate processing, the frequency spectrum that is obtained by FFT has been carried out interpolation carrying out FFT conversion voice signal before.If the crude sampling rate is SR, after sampling rate is reduced to original 1/RDC, carry out the FFT conversion that FFTLen is ordered, the power spectrum that obtains is used Inpl N point interpolation (that is: every adjacent amount FFT point interleaves Inpl_N-1 point) again, and the fundamental frequency that obtains so extracts resolution and is exactly: $\frac{\mathrm{SR}}{\mathrm{RD}{C}^{*}\mathrm{FFTLe}{n}^{*}\mathrm{Inpl}\_N}---\left(7\right)$ In present embodiment, original signal sampling rate SR=16000Hz is reduced to 1/4, that is: RDC=4, and the FFT conversion is counted and is FFTLen=512, interpolation Inpl_N=20 point.Extracting resolution like this is exactly 0.39Hz.And for the autocorrelation function extraction method of 384 sampled points of every frame, extraction resolution is 16000/N ²(N is the signal period corresponding sampling points), promptly approximately from 15Hz to 0.4Hz, obviously it has sufficiently high resolution in the fundamental frequency scope.

Adopt the advantage of foregoing invention to be: to extract this two kinds of methods in conjunction with time domain and frequency domain fundamental frequency, can carry out effective complementation: by frequency domain handle according to harmonic wave and size determine the fundamental frequency candidate, the centre does not have artificial settings such as thresholding, can reduce the uncertainty amount of calculating; On the dynamic programming function, considered time domain auto-correlation and frequency domain harmonic wave and these two and the closely-related important indicator of fundamental frequency simultaneously, make it maximum, guaranteed the continuity and the reliability of fundamental frequency; Random length is recalled the theory time-delay that then can reduce speech recognition system, and this requires more intense extremely important as call voice identification etc. to real-time.

For validity more of the present invention, disturb strong, fundamental frequency to extract two difficult voice and extracted 725 frame voiced sounds from sound channel, through artificial judgment, with tangible discontinuous fundamental frequency as identification error.Do not adopting before fundamental frequency candidate among the present invention adds DP algorithm, the identification error rate is 11.7%, is 0.4% after adopting.The accuracy rate of identification improves greatly.

In conjunction with foregoing description, the flow process that complete fundamental frequency extracts as shown in Figure 6: (1) signal segmentation, the voice signal of input at first is split into some frames, adjacent two frames have certain overlapping; Respectively every frame is carried out following processing then; (2) zero-crossing rate detects, and calculates average zero passage number of times, carries out the guestimate of pure and impure sound; The zero passage number of times is higher than given fault value and is judged as non-unvoiced frame, does not carry out fundamental frequency and extracts; (3) down-sampled rate had both improved fundamental frequency and had extracted resolution, can guarantee again not lose that 1250Hz is above that fundamental frequency is extracted significant frequency content; (4) pre-emphasis and windowing, the purpose that this one handles is to have reduced frequency alias.What adopt is the hamming window.Formula is: (5) FFT calculates and asks power spectrum, employing be 512 FFT (fast fourier transform).Use formula

S (f) | X (f) | ²The rated output spectrum; (6) interpolation in order to improve the precision that fundamental frequency extracts, is inserted Inpl_N=20 value between per two FFT point values of frequency spectrum, use be secondary Spline interpolation method; (7) calculate that harmonic wave adds up and, the formula that harmonic wave adds up is: (8) determine the fundamental frequency candidate, select several peaks from harmonic wave adds up spectrum, the fundamental frequency of peaking point correspondence, and the relative height of this peak value and maximum peak, that is: H _Per $H\left(f\right)=\underset{n=1}{\overset{\mathrm{HN}}{\mathrm{\Σ}}}{h}_{n}S\left(\mathrm{nf}\right)$ (9) candidate to a plurality of fundamental frequencies asks corresponding coefficient of autocorrelation, that is: R on time domain _PerFor having low H _PerValue or R _PerThe fundamental frequency candidate of value is filtered, to reduce the operand of DP; (10) ask the fundamental frequency path with DP algorithm, calculate the score of every paths with DP algorithm, the score formula is:

Score (i)=max{Score (i-1)-D (i, j) }+aR _Per(i)+bH _Per(i) also can be designated as Score (Dis, H _Per, R _Per); Wherein Dis represents the variation of fundamental frequency, R _PerThe expression coefficient of autocorrelation, H _PerThe relative height at expression peak; According to score, get several paths of top and note; (11) handle one section continuous unvoiced frame, then export as the result who extracts with the fundamental frequency path of optimum according to the score of DP algorithm; Otherwise returning zero-crossing rate detects; Handle whole input signals, enter normalization and be connected,, carry out normalization, can eliminate speaker's difference like this with average fundamental frequency to one section voice of handling; Connection is that the consonant part that does not have fundamental frequency is partly carried out level and smooth being connected with the vowel with fundamental frequency.

Description of drawings: Fig. 1. fundamental frequency extracts example as a result, and the left side is an isolated voice, and the right is a continuous speech.Fig. 2. the spectrogram that a unvoiced frame obtains through FFT result.At F ₀, 2F ₀, 3F ₀There is tangible peak at the place, but F ₀The peak at place does not have 2F ₀Obvious Fig. 3. the harmonic wave of the unvoiced frame figure that adds up.At F ₀Tangible maximum peak is arranged, together with also having other several peaks (shown in the circle) in addition, as candidate Fig. 4 of fundamental frequency. the autocorrelator trace figure of a unvoiced frame.At T ₀, 2T ₀3T ₀There is bigger peak value at the place, but T ₀The height and the 2T at peak, place ₀, 3T ₀Very near Fig. 5. the autocorrelator trace figure of a unvoiced frames.Except R[0] peak value thought is all lower, also do not have obvious periodic Fig. 6. and fundamental frequency extracts process flow diagram.

Embodiment:

A tone that distinguishing feature is exactly a Chinese of Chinese speech.The standard Chinese tone generally is divided into, two, three, the four tones of standard Chinese pronunciation, that is: high and level tone, rising tone, last sound, falling tone.Also have softly (zero tone) in addition.The female syllable of forming of identical sound has diverse meaning, corresponding different word or speech probably.(for example: rise, week, prison term, the departure date, stink smell.) therefore, tone has very important significance in the identification of Chinese language.

Tone is embodied in fundamental frequency F ₀The variation of track.What Fig. 1 represented on the left side is the tone and the fundamental frequency contrast of isolated voice, and the right is the contrast figure of continuous speech tone and fundamental frequency.And the fundamental frequency extracting method that the present invention proposes can provide pitch contour line comparatively accurately for speech recognition, is used for Tone recognition; Particularly can be with the fundamental frequency that extracts parameter as speech recognition system, or set up the tone model, to reflect phonetic feature more comprehensively, instruct the processing of language, improve the accuracy of identification.

The accurate extraction of pitch contour line also can be used as the important evidence of information Recognition such as syntactic structure in the continuous speech, word are read again, speaker's purpose; The tone of Chinese is the difficult point in the Chinese teaching, the pronunciation that the demonstration by fundamental curve can the interactively instruction of papil, thereby can be used in the teaching of standard Chinese.Though tone can be subjected to contextual influence, tone information more is not subjected to the influence of acoustic enviroment and passage, thereby from some angles, tone is significant for the reliability that improves the voice messaging processing.

Claims (7)

1. A high-precision and high-resolution fundamental frequency extraction method for speech recognition is characterized in that the step of fundamental frequency extraction is: (1) signal segmentation, the voice signal of input is first divided into several frames, and two adjacent frames There is a certain overlap; then the processing of the next frame is generated separately; (2) Zero-crossing rate detection, calculate the average number of zero-crossing times, and perform a rough estimation of unvoiced sounds; the number of zero-crossings is higher than a given threshold and is judged as non-voiced sound frame, without fundamental frequency extraction; (3) down-sampling rate, which not only improves the resolution of fundamental frequency extraction, but also ensures that the frequency components above 1250Hz are not lost for fundamental frequency extraction; (4) pre-emphasis and windowing, this The purpose of some processing is to reduce frequency aliasing. The hamming window is used, and the formula is:

(5) FFT calculation and calculation of power spectrum, using multi-point FFT (fast Fourier transform), using the formula to calculate the power spectrum: S(f)=|X(f)| ² (6) Interpolation, in order to improve the accuracy of fundamental frequency extraction, insert Inpl_N values between every two FFT point values of the spectrum, using the secondary Spline interpolation method; (7) Calculate the sum of harmonic accumulation, the formula for harmonic accumulation is: $h\left(f\right)=\underset{\mathrm{no}=1}{\overset{\mathrm{HN}}{\mathrm{\Σ}}}{h}_{\mathrm{no}}S\left(\mathrm{nf}\right)$ (8) Peak detection, determine the fundamental frequency candidate, select several peaks from the harmonic accumulation spectrum, find the fundamental frequency corresponding to the peak point, and the relative height between the peak and the maximum peak, that is _Hper : (9) for multiple fundamental The corresponding autocorrelation coefficient is calculated in the time domain for the candidate of frequency, that is, R _per ; the fundamental frequency candidates with too low H _per value or R _per value are filtered out to reduce the amount of calculation of DP; (10) use DP The algorithm finds the fundamental frequency path, and the DP algorithm is used to calculate the score of each path. The formula is: Score(i)=max{Score(i-1)-D(i, j)}+aR _per (i)+bH _per (i) can also be recorded as Score(Dis, H _per , R _per ); where Dis Indicates the change of the fundamental frequency, R _per indicates the autocorrelation coefficient, and H _per indicates the relative height of the peak; according to the score, take the top few paths and record them; (11) After processing a section of continuous voiced sound frames, then according to the score of the DP algorithm, the optimal fundamental frequency path is used as the extracted result output; otherwise, return to the zero-crossing rate detection; after processing all the input signals, enter normalization and continuous Line, for a piece of processed speech, use the average fundamental frequency to normalize, so that the difference between speakers can be eliminated; the connection is to smoothly connect the consonant part without the fundamental frequency and the vowel part with the fundamental frequency. 2. the high-precision and high-resolution fundamental frequency extraction method for speech recognition according to claim 1 is characterized in that combining the fundamental frequency analysis of time domain and frequency domain, extracting possible several fundamental frequency candidates, and adopting a The objective function combining the autocorrelation coefficient, spectrum analysis results and fundamental frequency variation is searched by dynamic programming algorithm. Before performing autocorrelation calculation, according to the spectrum analysis results used for speech recognition, FFT harmonic analysis method is used to analyze the speech Preliminary estimation of the fundamental frequency range, reducing the range of autocorrelation calculations, and has no effect on the accuracy of fundamental frequency extraction; the characteristics of the harmonic analysis method include downsampling rate, pre-emphasis and windowing, FFT calculation and in order to improve the resolution Perform steps such as secondary Spline interpolation. 3. the high-precision and high-resolution fundamental frequency extraction method for speech recognition according to claim 1 is characterized in that after completing the harmonic analysis, select several peaks from the harmonic accumulation spectrum, and ask the corresponding fundamental frequency of the peak point , and the relative height between the peak and the maximum peak, that is H _per , and then calculate the corresponding autocorrelation coefficient in the time domain for multiple fundamental frequency candidates, that is R _per , for those with too low H _per value or R _per value The base frequency candidates need to be filtered out. 4. according to claim 2, 3 described speech recognition uses high-precision high-resolution fundamental frequency extracting method, it is characterized in that: get the first few candidates of spectrum analysis result as fundamental frequency, then according to its basic fundamental frequency locus of human pronunciation Based on the principle that it is impossible to generate mutations, a dynamic programming method is used to determine the correct fundamental frequency sequence from these candidates. 5. the high-precision and high-resolution fundamental frequency extraction method for speech recognition according to claim 1, is characterized in that because when searching for an objective function, the fundamental frequency cannot produce a sudden change, and the autocorrelation accumulation value of the whole path is made as far as possible The maximum and the harmonic peak accumulation value are the largest, and the corresponding objective function is proposed: Score(i)=max{Score(i-1)-D(i, j)}+aR _per (i)+bH _per (i) Where Score(i) represents the score of the fundamental frequency path in the i-th frame, a, b are coefficients, D(i, j) Indicates the distance between the fundamental frequency p _i of the i-th frame and the j-th fundamental frequency candidate p _i of the i-1th frame: D(i,j)＝2 ^* |p _i -p _j |/(p _i +p _j ) 6. the high-precision and high-resolution fundamental frequency extraction method for speech recognition according to claim 1 is characterized in that in using the HMM based on continuous density, it is necessary to smooth the fundamental frequency, and use the direct connection method to extract the fundamental frequency without The consonant part with the base frequency is concatenated with the vowel part with the fundamental frequency and normalized with the average fundamental frequency to remove speaker influence. 7. according to claim 4 and 6 described high-precision high-resolution fundamental frequency extraction methods for speech recognition, it is characterized in that: according to the characteristics of non-voiced sound without obvious fundamental frequency, adopting indefinite length backtracking method, reducing non-voiced Interference to DP caused by voiced sound random fundamental frequency extraction results; the combination of zero-crossing rate detection and autocorrelation coefficient is also used as the basis for judgment, and then a zero-crossing rate threshold and autocorrelation coefficient threshold with better discrimination are selected according to the statistical results Value, the frame whose parameter is less than this threshold is judged as non-voiced sound, and does not participate in the DP algorithm to find the fundamental frequency contour line. When cutting out most of the non-voiced sound frames without fundamental frequency, these non-voiced sound noise classes are reduced by using the DP algorithm The effect of extracting the fundamental frequency reduces errors.

Images