80 results about "Speech enhancement algorithm" patented technology

The present invention helps to reduce the noise level and to enhance the quality of speech signals, in communications, computers, entertainment and other applications, where microphones and loudspeakers are involved. Additionally, the invention includes a new noise reduction and speech enhancement algorithm which is created based on the principles of human hearing mechanism. Further, the algorithm uses a new set of speech recognition parameters instead of just signal-to-noise ratio (“SNR”) as used in the prior art.

The invention provides a double-microphone speech enhancer, which comprises a double-microphone array module, a delay compensation module, a cepstrum field dereverberation module, a speech enhancement processing module and an output module, wherein the cepstrum field dereverberation module has the effect of dereverberation. The invention also provides a speech enhancement method based on the double-microphone speech enhancer. The invention adopts the cepstrum field dereverberation module to carry out beam formation and low-pass filtering, and can effectively eliminate the affection of room reverberation, and meanwhile, by using a speech enhancement algorithm, background noise can be further suppressed.

The invention discloses a voice enhancement method employing a combination of nesting-subarray-based post filtering and spectrum-subtraction and is suitable for indoor environment, comprising the enhancement of multi-channel voice signal in vehicle environment; as the problems of unstable the broadband of the voice signals, the inconsistent frequency response of the microphone-array-based multi-channel voice-enhancement method to the voice signal and the correlation among all-channel noise in actual noise-field environment are considered, by utilizing the microphone array nested by the subarrays with different spacing, the voice signals are collected; and the voice signals formed by subarray beams are divided into a high-frequency section and a low-frequency section, different voice-enhancement algorism are adopted for carrying out the treatment; all the advantages are complementary with each other, thus improving the effect of voice enhancement.

The invention provides a voice enhancement method. The method comprises the following steps of: judging whether the current frame has pure noise by using a judgment device; if the current frame has pure noise and a plurality of previous frames of the current frame have pure noise, improving frequency domain signals by using a voice enhancement algorithm of an improved spectral subtraction method, otherwise, improving the frequency domain signals by using an enhancement algorithm of a voice generating model; and transforming the processed frequency domain signals to a time domain, performing de-emphasis processing and acquiring output signals. The invention also provides a device using the method. The voice enhancement method greatly improves the attenuation of residual noise, and ensures the voice intelligibility.

The invention discloses a backend i-vector enhancement method for a speaker recognition system. The method is based on a deep neural network; and a backend i-vector regression model for the speaker recognition system is built on the basis of the application of the deep neural network in speech enhancement, and a backend feature processor applicable to the speaker recognition system is obtained. Compared with a conventional front-end speech enhancement algorithm, the backend i-vector enhancement method of the invention can improve the anti-noise performance of the speaker recognition system and optimize the structure model of the speaker recognition system, so that the practicability of the speaker recognition system in a noise environment can be effectively improved.

The invention relates to noise identification based on CNN and a voice enhancement model--SFTRLS-CNN combining the CNN and smooth and fast recursive least squares (SFTRLS). Firstly, 648-dimensional characteristics of MFCC with a noise in a noise frequency band and the like are extracted and enter into a first kind of trained convolutional neural network so as to identify the environment type of the noise. Then, an extracted audio frequency characteristic, a signal to noise ratio and a noise type value are used to form 658-dimensional characteristics. A second kind of convolutional neural network is adopted to be adaptively matched to an SFTRLS algorithm so as to carry out the optimum forgetting factor of voice enhancement. Finally, through the smooth and fast recursive least squares, the noise reduction processing of various environments is realized. By using the algorithm, the enhancement model is suitable for the different noise environments and an adaptive capability is increased. Compared to a traditional SFTRLS, the algorithm has a better voice quality evaluation index value.

The invention discloses an end-to-end voice enhancement method based on generation of a countermeasure network. The method comprises the following steps of directly inputting a noisy voice signal intoa pre-trained deep neural network for signal processing and outputting an enhanced voice signal, wherein the depth neural network is obtained through training by the following steps of S1, preliminarily training to generate a countermeasure network, wherein the generation of the countermeasure network comprises two deep neural networks: a generator G and a discriminator D; S2, performing knowledge distillation on the simulated noisy voice through a traditional statistical speech enhancement algorithm, and then training to generate a countermeasure network again; S3, performing fine adjustmenton the generator G obtained through training through the real noisy voice; and S4, outputting the generator G trained in the above step as a final deep neural network for voice enhancement processing.

A telephone includes at least two microphones and a circuit for processing audio signals coupled to the microphones. The circuit processes the signals, in part, by providing at least one statistic representing maximum normalized cross-correlation of the signals from the microphones, doaEst, dirGain, or diffGain and comparing the at least one statistic with a threshold for that statistic. At least one of noise reduction and speech enhancement is controlled by an indication of near-field sounds in accordance with the comparison. Indication of near-field speech can be further enhanced by combining statistics, including a statistic representing inter-microphone level difference, each of which have their own threshold. dirGain and diffGain are derived from signals incident upon the microphones such that the desired near-field signal is not suppressed.

The invention discloses a self-adaptive spectral subtraction real-time speech enhancement method. The method comprises: establishing a dynamic threshold including discrimination of voice and no voice of noise voice, providing noise spectrum time varying update principles according to the dynamic threshold; making full use of correlation extraction information among adjacent frames, and realizing a pure voice spectrum smooth estimation method. Aimed at a practical problem that voice signals are difficult to extract in unstable noise and strong background noise, the self-adaptive spectral subtraction speech enhancement method is provided. The method uses a rapid tracking noise algorithm to perform smooth update on the unstable noise frame by frame, and can preferably estimate noise spectrums. The algorithm can effectively restrain background noise, and improves voice quality and intelligibility after noise reduction. The method is low in calculation cost, and is easy to realize, and has good real-time property. The method provides a new approach for denoising of strong background noise and detection of weak signals.

The invention discloses a speech enhancement method based on Gaussian mixture model (GMM) noise estimation, wherein the GMM is used for estimating background noise and a spectral subtraction coefficient, spectral subtraction is conducted on noisy speech, and pure speech is recovered. Firstly, the noisy speech is preprocessed so as to obtain the amplitude and phase of the noisy speech, the amplitude is used for noise estimation and spectral subtraction, and the phase is used for recovering a time-domain signal; then, the GMM is used for estimating noise parameters and pure speech cepstrum characteristics from the noisy speech in real time, and the spectral subtraction coefficient is calculated according to the estimated pure speech cepstrum characteristics; finally, spectral subtraction is conducted on the frequency spectrum of the noisy speech, the time-domain signal is recovered, and enhanced speech is obtained according to an overlap-add method. According to the speech enhancement method, the capability of the speech enhancement algorithm to track non-stationary noise can be improved remarkably.

The invention discloses an objective assessment method based on voice enhancement algorithm subjective assessment. The method is a new objective combination method proposed by utilizing multi-linear regression analysis based on five voice enhancement objective assessment methods as follows: PESQ, LLR, CEP, fwSNRseg and fwSNRsegVar. An objective score for assessing the quality of a voice signal, an objective score for assessing the quality of background noise and an objective score for assessing overall quality are obtained by the combinations of different weighting coefficients and five objective methods. The subjective and objective correlation coefficients of the method are superior to a plurality of other traditional objective algorithms, the standard deviation of error is lower, and the assessment of performance of the voice enhancement algorithm is easier, therefore, the invention has great practical value.

The invention discloses a prior signal-to-noise ratio estimating method based on an MMSE error criterion and used for voice enhancement, and belongs to the technical field of voice signal processing. Aimed at the prior signal-to-noise ratio estimating problem in the voice enhancement technology, the method comprises the steps of: firstly carrying out preliminary estimation on a prior signal-to-noise ratio of noised voices based on the MMSE error criterion, carrying out Wiener filtering calculation on an obtained prior signal-to-noise ratio estimated value to obtain a first system gain factor, carrying out calculation on the first system gain factor an amplitude spectrum value of the noised voices to obtain a voice power spectrum estimated value, then utilizing the obtained voice power spectrum estimated value and a power spectrum estimated value of noise to carry out estimation once again, and obtaining a final prior signal-to-noise ratio estimated value. The prior signal-to-noise ratio estimated value is substituted into a subsequent voice enhancing step for processing, and de-noised estimated voice clearing signals are obtained. The prior signal-to-noise ratio estimating method based on the MMSE error criterion can effectively inhibit background noise components in estimated cleared voices, and excessive damages to the cleared voice components are avoided, so that the hearing quality of the estimated cleared voices is improved, and the performance of a voice enhancement algorithm is improved.

The invention discloses a method for eliminating nonuniform noise in speech collected by radar. The method comprises the steps that data windowing is carried out on a noisy radar speech signal y(n); according to a given decomposition scale, discrete wavelet transform (DWT) decomposition is carried out on the speech signal; wavelet entropy (WE) calculation is carried out; noise estimation is carried out based on dynamically-updated wavelet entropy (WE); and according to the updated noise estimation, Wiener filtering is used to enhance the speech signal. According to the invention, dynamic advantages of wavelet entropy noise estimation and advantages of Wiener filtering are combined to form a wavelet entropy Wiener filtering speech enhancement algorithm, thus nonuniform colored noise of radar speech can be effectively filtered out.

The invention provides a directional speech enhancement method based on a small microphone array, which comprises the following steps of: 1) utilizing two fully directional microphones to acquire a sound signal; 2) utilizing a self-adaptive null-forming algorithm to perform data processing on the acquired sound signal to obtain a delay substraction signal x(t) and a signal z(t) after the self-adaptive filtering; 3) processing the x(t) and the z(t) to form speech spectrums, namely X (omega) and Z(omega) respectively; 4) calculating preliminary gain G'(omega) by utilizing a single-channel speech enhancement method according to the X (omega) and the Z(omega), and calculating the existence probability P(omega) of a target signal according to the X (omega) and the Z(omega); 5) utilizing the existence probability P(omega) of the target signal to amend the preliminary gain G'(omega) to obtain final gain G(omega), wherein the G(omega) is equal to (the G'(omega))Gm, and Gm is a preset minimum gain value; and 6) utilizing the final gain G(omega) to enhance the signal z(t) after the self-adaptive filtering to obtain a final enhanced speech signal r(t). The method can realize a directional speech enhancement algorithm in small volume, can obtain noise suppression to a greater extent, and improve the signal-to-noise ratio.

The invention discloses an improved self-coding neural network voice enhancement algorithm and aims at problems in the voice enhancement effect existing in traditional voice enhancement algorithms such as a spectral subtraction algorithm and a Wiener filtering method, e.g., the poor filtering effect for the non-stable noise, residual of music noise after enhancement and the poor generalization effect for noise types and the signal to noise ratio. According to the algorithm, a three-level junction of the self-coding neural network is enhanced to be a five-level junction, the neuron numbers of the correspondingly levels are respectively 256, 128, 64, 128 and 256, 100 types of noise are added to pure audios according to signal to noise ratios of -5dB, 0dB, 5dB, 10dB and 15dB, mass data sets are constructed to train the network, the excellent voice enhancement effect can be realized after the network model is trained, as the training data is big, the excellent generalization effect for thenoise types and the signal to noise ratio is realized.

The invention discloses a speech enhancement method for generating an adversarial network based on a two-dimensional spectrogram and conditions. The method comprises the following steps of forming thetwo-dimensional spectrogram through a plurality of frame spectrums obtained by performing short-time Fourier transform on a speech signal, generating input characteristics of the adversarial networkby serving the two-dimensional spectrogram as the condition, and generating a network G through the mutual adversarial training of a generating network G and a discrimination network D. In the testingprocess, the two-dimensional spectrogram of the noisy speech is extracted, and the G network obtained in the training stage directly maps the noisy speech spectrogram into an enhanced spectrogram, thereby realizing speech enhancement. Through the speech enhancement algorithm for generating the adversarial network based on the spectrogram and the conditions disclosed by the invention, the perception quality of the enhanced speech is greatly improved, and the algorithm has good generalization performance and stronger robustness.

The invention relates to a voice enhancement method for use in the voice identification process of an electric wheelchair. According to the enhancement method, enhancement is performed by using a voice enhancement algorithm combining an audio masking effect and a priori signal-to-noise ratio, so that the problem of low identification rate of intelligent wheelchair robust voice identification in a noise environment is solved. The method has high generality to a specific human voice identification system. By applying the method to voice identification in an electric wheelchair voice control system, the identification rate of the system in the noise environment is increased, accurate control of the electric wheelchair is realized, and voice interaction between a user and the electric wheelchair is realized.

The invention discloses a method for enhancing voice based on signal to noise ratio soft masking. The method comprises: establishing noise power spectrum updates of a sub-band time varying coefficient, using different threshold update smooth spectrums for different frequency points, enhancing voice and restraining noise; determining a posterior signal to noise ratio from a noise power spectrum, performing iterative computation to obtain a prior signal to noise ratio of a present frame according to the posterior signal to noise ratio and the prior signal to noise ratio of a previous frame, and obtaining that each frequency point is in a masking region or in a target signal region according to values of the prior signal to noise ratio. A masking value is calculated by probability distribution obtained by hypothesis testing. The method uses correlation of adjacent frames to extract information to realize enhancement of voice spectrum smooth iteration estimation. For non-stationary noise and strong background noise, a voice enhancement algorithm based on signal to noise ratio soft masking is provided. A rapid tracking noise algorithm performs smooth update on the non-stationary noise frame by frame, preferably estimating a noise spectrum. The algorithm can effectively restrain background noise and improve voice quality and speech intelligibility after denoising.

The invention discloses a speech enhancement method and system. The method comprises the steps that a constructed speech enhancement network model comprises two parallel modules: a prediction adaptiveweight module judges a signal-to-noise ratio according to an input feature so as to adjust the proportion of speech distortion and residual noise through weight; and a prediction time-frequency maskmodule estimates a time-frequency mask for suppressing noise according to the input features. The proportion of speech distortion and residual noise in the enhanced speech can be adaptively adjusted according to the signal-to-noise ratio through the training network, and the trained network model is used for an actual noise reduction task to obtain an enhanced speech signal. According to the method, the neural network is used for adaptively adjusting and enhancing voice distortion and residual noise in the voice to obtain a better voice enhancement effect, different adaptive weight ranges canbe trained according to different task requirements, and a voice enhancement algorithm more suitable for related tasks is obtained.

The invention provides a voice enhancement processing method and a device. The method comprises the steps that voice information from terminal equipment is acquired, wherein voice enhancement auxiliary information is carried in the voice information; if voice enhancement processing is required to be performed on the voice information by judging according to equipment identification of the terminal equipment, a corresponding voice enhancement algorithm is acquired from a plurality of local voice enhancement algorithms according to the voice enhancement auxiliary information; and the voice enhancement processing is performed on the voice information according to the acquired voice enhancement algorithm. With the adoption of the technical scheme, a voice enhancement processing procedure can be more pertinent, and an unnecessary computation burden of a server is reduced under the condition that the voice enhancement quality is ensured.

The embodiment of the invention provides an adversarial sample attack defense method and device based on a voice enhancement algorithm. The method comprises the steps of obtaining a to-be-identified voice sample and spectrum features of the to-be-identified voice sample; according to the spectrum characteristics of the to-be-identified voice sample, calculating a noise spectrum of the to-be-identified voice sample through a preset algorithm, and denoising the to-be-identified voice sample by using the estimated noise spectrum obtained through calculation to obtain a denoised voice sample, wherein the algorithm comprises a spectral subtraction method based on continuous minimum tracking and a logarithm minimum mean square error algorithm MMSE algorithm combined with speech existence probability; and recognizing the denoised voice sample through a pre-trained voice recognition model to obtain a recognition result. Therefore, after the to-be-identified voice sample is obtained and de-noising processing is carried out on the to-be-identified voice sample, the de-noised voice sample is identified, so that the voice identification accuracy is improved and the efficiency of defending against the attack of the countermeasure sample is improved.

The invention discloses a hearing-aid multichannel voice enhancing algorithm based on iterative Wiener filtering. The algorithm has the following steps of dividing a noisy signal into a plurality of frames; performing noise estimation on each frame of noisy signal; calculating a posterior signal-to-noise ratio; estimating a prior signal-to-noise ratio by using the posterior signal-to-noise ratio;calculating the noise reduction attenuation gain required for noise reduction according to the posterior signal-to-noise ratio; acting the noise reduction attenuation gain onto the noisy signal so asto obtain an enhanced voice signal; using the obtained enhanced voice signal as given enhanced voice; continuously and repeatedly performing the front steps on the next frame of voice signals. The voice enhancing algorithm provided by the invention has the advantages that the algorithm complexity is low; the noise reduction effect is good; the hearing ability compensation is ensured; meanwhile, the voice recognition ability of a hearing-aid wearer can be improved.

The invention discloses a short-wave channel speech anti-fading auxiliary enhancement method based on a convolutional neural network, belongs to the technical field of communication, and particularly relates to a short-wave fading resistant speech enhancement auxiliary method. Firstly, an applicable short-wave speech communication model is defined; after a transmitting terminal obtains a speech signal sample, background environment noise is eliminated by using an existing speech enhancement technology, then SSB modulation is carried out, up-conversion is carried out to a short-wave frequency band for transmission, a transmitting signal reaches a receiver at a far end through a short-wave channel, and speech enhancement is carried out on the received signal after down-conversion and SSB demodulation, so that the purpose of the invention is achieved. The anti-fading convolutional neural network can be used for assisting most speech enhancement algorithms based on speech feature extraction and further improving the quality of short-wave received speech signals, for example, the anti-fading convolutional neural network can be combined with a spectral subtraction method, a method based on a statistical model, an NMF algorithm and the like described in the background technology.

The invention discloses a speech enhancement algorithm based on an attention mechanism. A neural network speech enhancement model based on the attention mechanism is constructed, and comprises three components of a neural network based on the attention mechanism, a standard deep loop neural network and a time-frequency masking layer. At each time step, the model performs attention mechanism calculation on an incoming frame at current time and speech frames of an entire segmentto obtain feature vector expression corresponding to the current time step. The model input is obtained by splicing thecurrent time step feature vector with the current speech frame, and the current input is encoded by the standard deep loop neural network to obtain a predicted value of time-frequency masking. The predicted value of the time-frequency masking is multiplied by the mixed speech step-by-step to obtain an enhanced speech segment. The algorithm models a speech enhancement problem from the perspectiveof improving the generalization performance of the model, and can effectively solve the speech enhancement problem in a noise scene which does not appear in the training.

The invention provides a single-channel speech enhancement algorithm based on a compensation phase spectrum. The single-channel speech enhancement algorithm comprises the following steps: preprocessing a noisy speech signal, and framing and windowing the noisy speech signal; carrying out fourier transform; dividing a critical frequency band by using an ERB scale; calculating the value of the segmented signal-to-noise ratio; calculating a new compensation factor, and obtaining a primarily enhanced voice complex frequency spectrum through power spectrum subtraction; performing additive calculation on the phase spectrum compensation function and the primarily enhanced voice complex frequency spectrum to obtain a compensated complex frequency spectrum; solving a phase angle of the compensatedcomplex frequency spectrum to obtain a compensated phase spectrum; and overlapping and adding the speech amplitude spectrum after basic spectrum subtraction and the compensation phase spectrum obtained in the seventh step, and then carrying out inverse Fourier transform to obtain an enhanced speech signal. The invention verifies that the improved compensation factor not only has an effect on steady-state noise, but also is more beneficial to the denoising effect of unsteady-state noise, and the speech enhancement algorithm provided by the invention is relatively wide and effective in applicable noise environment.

The invention provides a method used for improving the speech enhancement algorithm performance of a microphone array. The method comprises the steps that a speech detector detects whether a primary channel is currently a speech-time HSP or a non-speech-time NSP; in NSP, a first filter adaptively adjusts the coefficient of the first filter based on an input signal, so that the energy of an output signal of the first filter is minimized; in HSP, the coefficient of the first filter is locked without adjustment, and the input signal is directly filtered based on the filter coefficient set in NSP; the current output signal of the first filter is used the input of the second filter; self-adaptive filtering is carried out on a primary channel signal; and by adaptively adjusting the coefficient of the second filter, the energy of the output signal of the second filter is minimized to acquire anti-crosstalk and noise-canceled speech output. According to the invention, the speech enhancement effect can be improved to a certain extent.

The invention discloses a speech enhancement method applied to an ultra-short wave radio station. The speech enhancement method include that a received speech signal is preprocessed, and then an enhanced speech signal is obtained by using a maximum posterior probability algorithm; in a calculation process by using the maximum posterior probability algorithm, the a priori SNR parameter is optimizedby an adaptive filter. The method combines the characteristics of the ultrashort wave radio station, combines the ultrashort wave radio station and the speech enhancement technology more effectively,accurately estimates the noise, improves the speech enhancement effect, and optimizes the speech enhancement algorithm to significantly improve the SNR.

A Dynamic Noise Compensation (DNC) telephone speech enhancement algorithm addresses the issue of environment noise on the listener end of a telephone call. A single microphone proximal to the listener provides a sample of near end ambient noise level and of near end speech. A Voice Activity Detector (VAD) detects the presence of near end (listener) speech. The DNC algorithm adjusts the far end incoming speech level based on the near end ambient noise and the VAD ensures that the near end listener speech does not effect the incoming speech level adjustment.

The invention discloses a voice recognition method applied to ground-air communication. The voice recognition method comprises the steps: a ground-air communication triphone acoustic model is established; through an improved maximum posterior probability speech enhancement algorithm, a received to-be-recognized ground-air communication voice signal is subjected to speech enhancement and backgroundnoise removal processing; the processed to-be-recognized ground-air communication voice signal is input into the ground-air communication triphone acoustic model to be recognized, speech command textand keyword text of a controller and a pilot are recognized, and when it is recognized that the speech command text of the controller and the speech command text of the pilot are inconsistent, alarmreminding is conducted; and the recognized keyword text is detected through a keyword detection model, and when preset words are detected, alarm reminding is conducted. According to the voice recognition method, voice commands between the controller and the pilot can be recognized and compared, the sensitive words can further be detected, alarm reminding is conducted, and the voice recognition rate can be increased.