Method and apparatus for voice spoof detection in noisy environments

Abstract

The present application provides a method for voice spoof detection in a noisy environment. The method comprises: receiving a voice input of a user; receiving an environmental noise measurement value corresponding to the voice input of the user; authenticating the user according to the voice input; determining a noise parameter according to the environmental noise measurement value; determining an actual value of a Lombard effect related feature of user voice according to the voice input of the user; determining a voice spoof probability by comparing the actual value of the Lombard effect related feature with a Lombard effect related feature template of the authenticated user corresponding to the determined noise parameter.

Description

Technical Field

[0001] This invention relates to speech spoofing detection. More specifically, it relates to a method for performing speech spoofing detection in noisy environments, thereby increasing protection and improving accuracy. The invention also relates to an apparatus for performing speech spoofing detection in noisy environments, thereby increasing protection and improving accuracy. Background Technology

[0002] Most electronic devices employ biometric security measurements, or biometric authentication, to allow use of the device or to perform any operation within it. Voice authentication is one type of biometric authentication, widely used in modern voice assistants, voice control systems, and so on. Voice authentication matches spoken words against a biometric profile to grant access. Many access services' voice authentication systems are equipped with an authentication / verification subsystem (AVS), so these systems may have limitations for each individual. As known in Alexa, a typical voice verification system is coupled with an activation phrase detector and uses a wake-up phrase to authenticate the user. For example, a wake-up phrase is a voice pattern that matches a pre-recorded template. These types of voice authentication systems can be vulnerable to spoofing attacks, such as voice playback and speech synthesis.

[0003] Currently, several known voice anti-spoofing techniques exist, including detecting artifacts in recorded / replayed audio and detecting breaks in the continuity of speech background. In voice authentication systems, existing voice anti-spoofing techniques heavily rely on noise-dominated environmental parameters. High noise levels in automotive, industrial, and open-space applications provide edge anti-spoofing engines, allowing attackers to leverage noise injection to launch continuous attacks on voice authentication systems. Furthermore, these voice authentication systems incorporate several complementary techniques to establish a viable level of protection. A key issue with existing voice anti-spoofing systems is their low noise tolerance. Spoofing detection errors can increase with high noise levels, rendering the security and user experience of the voice authentication system unacceptable.

[0004] Therefore, the present invention aims to improve the performance of existing systems or technologies in voice spoofing detection. Summary of the Invention

[0005] One object of the present invention is to provide a method and apparatus for voice spoofing detection in noisy environments, thereby enhancing protection and improving accuracy, while avoiding one or more disadvantages of existing methods.

[0006] This objective is achieved through the features of the independent claim. Other implementations are apparent in the dependent claims, the specification, and the drawings.

[0007] This invention provides a method for detecting speech spoofing in noisy environments. This invention also provides an apparatus for detecting speech spoofing in noisy environments.

[0008] According to a first aspect, a method for detecting speech spoofing in a noisy environment is provided. The method includes: receiving a user's voice input. The method includes: receiving an ambient noise measurement corresponding to the user's voice input. The method includes: authenticating the user based on the voice input. The method includes: determining a noise parameter based on the ambient noise measurement. The method includes: determining an actual value of a Lombard effect-related feature of the user's voice based on the user's voice input. The method includes: determining a speech spoofing probability by comparing the actual value of the Lombard effect-related feature with a Lombard effect-related feature template of the authenticated user corresponding to the determined noise parameter.

[0009] This method offers better protection against voice spoofing attacks in noisy environments. It also prevents replay, imitation, and speech synthesis attacks. The method can include additional independent authentication factors to authenticate users, thereby ensuring system security. It is relatively lightweight. Even in very noisy environments, it improves the accuracy of user voice input by increasing the probability of voice spoofing. Furthermore, it introduces a complementary technique that utilizes noise for voice spoofing detection, significantly enhancing voice anti-spoofing capabilities in noisy environments.

[0010] Optionally, the Lombard effect-related features include changes in the user's speech characteristics in the time domain, spectral domain, and power spectral domain under normal conditions and in the noisy environment, including changes in speech fundamental frequency, energy distribution between high, mid, and low frequency bands, sound intensity, vowel duration, spectral tilt, format center frequency, keyword duration, and lung volume.

[0011] Optionally, the noise parameters include noise type and noise level.

[0012] Optionally, the Lombard effect-related feature template includes preset values ​​of the Lombard effect-related features of the user's speech under different noise parameters.

[0013] Optionally, the preset value of the Lombard effect-related feature includes the average value of the Lombard effect-related feature collected for the user's speech under different noise parameters during the training period.

[0014] Optionally, based on the Lombard effect-related feature template corresponding to the determined noise parameters of the authenticated user, the voice spoofing probability is determined as a function of the difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature.

[0015] Optionally, the Lombard effect-related feature template further includes weights for the Lombard effect-related features, and the speech deception probability is determined as a function of the weighted difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, wherein the weighted difference is obtained through the weights.

[0016] Optionally, the method further includes: verifying the user's authentication based on the voice spoofing probability.

[0017] Optionally, the method further includes: in response to verifying the user's authentication, updating the user's Lombard effect-related feature template using the actual value of the Lombard effect-related feature.

[0018] According to a second aspect, an apparatus for detecting speech spoofing in a noisy environment is provided. The apparatus includes: an input module for receiving a user's voice input and an ambient noise measurement corresponding to the user's voice input; a voice verification module for authenticating the user based on the voice input; a noise retrieval unit for determining noise parameters based on the ambient noise measurement; a Lombard feature retrieval unit for determining actual values ​​of Lombard effect-related features based on the user's voice input; and a spoofing detector for determining a speech spoofing probability by comparing the actual values ​​of the Lombard effect-related features with a Lombard effect-related feature template of the authenticated user corresponding to the determined noise parameters.

[0019] The aforementioned device provides better protection against voice spoofing attacks in noisy environments. It also prevents replay, imitation, and speech synthesis attacks. The device can include additional independent authentication factors to authenticate users, thereby ensuring system security. It is relatively lightweight. Even in very noisy environments, the device improves the accuracy of user voice input by increasing the probability of voice spoofing. The device introduces a complementary technique that utilizes noise for voice spoofing detection, significantly enhancing voice anti-spoofing capabilities in noisy environments.

[0020] Optionally, the Lombard feature retrieval device is used to determine the actual values ​​of the Lombard effect-related features by estimating the changes in the features of the user's speech in the time domain, spectral domain, and power spectral domain under normal conditions and in a noisy environment, wherein the changes include changes in the speech fundamental frequency, energy distribution between high, mid, and low frequency bands, sound intensity, vowel duration, spectral tilt, format center frequency, keyword duration, and lung volume.

[0021] Optionally, the noise retriever is used to determine the noise type and noise level.

[0022] Optionally, the Lombard effect-related feature template includes preset values ​​of the Lombard effect-related features of the user's speech under different noise parameters.

[0023] Optionally, the preset value of the Lombard effect-related feature includes the average value of the Lombard effect-related feature collected for the user's speech under different noise parameters during the training period.

[0024] Optionally, the spoofing detector is used to determine the speech spoofing probability as a function of the difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, based on the Lombard effect-related feature template corresponding to the determined noise parameters of the authenticated user.

[0025] Optionally, the Lombard effect-related feature template further includes weights for the Lombard effect-related features, and the spoofing detector is used to determine the speech spoofing probability as a function of the weighted difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, wherein the weighted difference is obtained through the weights.

[0026] Optionally, the device further includes a decision module for verifying the user's authentication based on the voice spoofing probability.

[0027] Optionally, the device further includes a data acquisition unit, which is used to acquire for each user the features of the user's speech under normal conditions, noise parameters, the actual values ​​of the Lombard effect-related features of the user's speech under the noise parameters, and the Lombard effect-related template.

[0028] Optionally, the apparatus further includes a learning engine configured to: update the user's Lombard effect-related feature template using the actual values ​​of the Lombard effect-related features in response to the decision module and / or learning trigger verifying the user's authentication.

[0029] Therefore, compared with existing technologies, the above-described methods and apparatus can better prevent voice spoofing attacks in noisy environments. Even in very noisy environments, the above-described methods and apparatus can improve the accuracy of users' voice input by increasing the probability of voice spoofing.

[0030] These and other aspects of the invention will be apparent from one or more implementations described below. Attached Figure Description

[0031] The implementation of the present invention will now be described by way of example only, with reference to the accompanying drawings, in which:

[0032] Figure 1 This is a block diagram of an apparatus for detecting voice spoofing in a noisy environment, provided in one implementation of the present invention;

[0033] Figure 2 This is a block diagram of an apparatus for detecting voice spoofing in a noisy environment, provided in one implementation of the present invention;

[0034] Figure 3 This is a graphical representation of the deception probability provided by one implementation of the present invention;

[0035] Figure 4 This is a system architecture for a device used for voice spoofing detection in a noisy environment, provided by one implementation of the present invention;

[0036] Figure 5A and Figure 5B This is a flowchart of a method for detecting speech spoofing in a noisy environment, provided by one implementation of the present invention. Detailed Implementation

[0037] The present invention provides a method and an apparatus for detecting speech spoofing in noisy environments, thereby enhancing protection and improving accuracy.

[0038] To enable those skilled in the art to more easily understand the present invention, the following implementation of the present invention is described with reference to the accompanying drawings.

[0039] In the description of the invention, the claims, and the accompanying drawings, the terms "first," "second," "third," and "fourth" (if any) are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that these terms are interchangeable where appropriate, and therefore, the embodiments of the invention described herein can be implemented in orders other than those described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to the explicitly listed steps or units, but may include other steps or units not explicitly listed or inherent to the aforementioned process, method, product, or apparatus.

[0040] Figure 1 This is a block diagram of a device 102 for detecting voice spoofing in a noisy environment, provided in one implementation of the present invention. Device 102 includes: an input module 104 for receiving a user's voice input and an ambient noise measurement corresponding to the user's voice input; a voice verification module 106 for authenticating the user based on the voice input; a noise retrieval unit 108 for determining noise parameters based on the ambient noise measurement; a Lombard feature retrieval unit 110 for determining the actual values ​​of Lombard effect-related features based on the user's voice input; and a spoofing detector 112 for determining the probability of voice spoofing by comparing the actual values ​​of the Lombard effect-related features with the Lombard effect-related feature template of the authenticated user corresponding to the determined noise parameters.

[0041] Device 102 provides better protection against voice spoofing attacks in noisy environments. Device 102 can prevent replay, imitation, and speech synthesis attacks. Device 102 can include additional independent authentication factors to authenticate users, thereby ensuring system security. Device 102 is relatively lightweight. Even in very noisy environments, device 102 can improve the accuracy of user voice input by increasing the probability of voice spoofing. Device 102 introduces a complementary technique that utilizes noise for voice spoofing detection, significantly improving voice anti-spoofing capabilities in noisy environments.

[0042] The Lombard effect, or Lombard reflex, refers to the unconscious increase in volume by a speaker in noisy environments to improve the inaudibility of their speech. The Lombard feature retrieval unit 110 is used to determine the actual values ​​of Lombard effect-related features based on the Lombard effect. The Lombard effect includes loudness and other acoustic features such as pitch, rate, syllable duration, etc. The Lombard effect can preserve the speaker's signal-to-noise ratio (SNR).

[0043] Device 102 can transform a user's raw speech input into Lombard-adjusted speech. Optionally, Lombard-adjusted speech may include any of the following: increasing the fundamental frequency, shifting energy from the low-frequency band to the mid- or high-frequency band, increasing sound intensity, prolonging vowel duration, spectral tilting, shifting the format center frequency, extending the duration of content words to a greater extent, or using greater lung capacity. Optionally, the Lombard-adjusted speech may vary depending on the type and level of noise.

[0044] Optionally, the Lombard feature retriever 110 is used to determine the actual values ​​of Lombard effect-related features by estimating the changes in the features of the user's speech in the time domain, spectral domain, and power spectral domain under normal conditions and in noisy environments, wherein these changes include changes in the speech fundamental frequency, energy distribution between high, mid, and low frequency bands, sound intensity, vowel duration, spectral tilt, format center frequency, keyword duration, and lung volume.

[0045] Optionally, the noise retrieval unit 108 is used to determine the noise type and noise level. The noise retrieval unit 108 can determine the noise type and level perceptual Lombard characteristics for each user.

[0046] Optionally, the Lombard effect-related feature template includes preset values ​​of Lombard effect-related features of the user's speech under different noise parameters.

[0047] Optionally, the preset values ​​of the Lombard effect-related features include the average values ​​of the Lombard effect-related features collected for user speech under different noise parameters during the training period.

[0048] Optionally, the spoofing detector 112 is used to determine the speech spoofing probability as a function of the difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, based on the Lombard effect-related feature template of the authenticated user corresponding to the determined noise parameters.

[0049] Optionally, the Lombard effect-related feature template also includes weights for the Lombard effect-related features. The deception detector 112 uses a function to determine the speech deception probability as a weighted difference between the actual value of the Lombard effect-related feature and a preset value of the Lombard effect-related feature, wherein the weighted difference is obtained through weights.

[0050] Figure 2This is a block diagram of a device 202 for detecting voice spoofing in a noisy environment, provided in one implementation of the present invention. Device 202 includes an input module 204, a voice processing module 206, a noise classification and measurement module 208, a voice authentication module 210, and a voice anomaly detection module 212. Optionally, device 202 further includes a decision module 214, which verifies the user's authentication based on the probability of voice spoofing. Device 202 can be placed in a vehicle. Optionally, the vehicle is a car. Input module 204 is used to monitor either the user's voice or ambient noise. The user's voice can be at least one of conversation or command. Ambient noise can be any noise emitted by tires, engine, air conditioning, road surface, in-car music, etc. Optionally, input module 204 is a microphone used to monitor either the user's voice or ambient noise. Optionally, input module 204 acquires the user's voice input.

[0051] The speech processing module 206 processes either the user's speech input or ambient noise using conventional feature methods and Lombard feature methods. Conventional feature methods process the user's speech input to convert it into an algorithm. Optionally, these algorithms are used to train a learning engine and / or to enable the learning engine to perform one or more operations. Lombard feature methods process the user's speech input to enhance it. The noise classification and measurement module 208 determines the noise type (Perceptual Linear Prediction, PLP) and noise level of either the user's speech input or ambient noise. The Lombard feature method in the speech processing module 206 determines Lombard speech parameters based on the noise type and noise level of either the user's speech input or ambient noise. The determined Lombard speech parameters can be compared with the actual Lombard parameters (ALP) retrieved from the user's speech input. Optionally, the determined Lombard speech variations can be learned during training of the learning engine.

[0052] Voice authentication module 210 is used to authenticate a user based on the user's voice input. Optionally, the learning engine can only perform one or more operations based on the user's voice input if the voice authentication module 210 authenticates the user. Voice authentication module 210 may include a general template that includes one or more controls for performing one or more operations based on the user's voice input. Optionally, the general template includes one or more models and templates and configurations for one or more users. The learning engine can create one or more models for one or more users. Voice anomaly detection module 212 is used to determine the probability of spoofing in the user's voice input. The probability of spoofing can be calculated as a function of the difference between the actual Lombard parameter (ALP) and the determined Lombard speech parameter. Optionally, the probability of spoofing is calculated as a function of the difference between ALP and the measured parameter based on the noise type PLP, i.e., P = f( / PLP = –ALP / ).

[0053] Decision module 214 is used to verify user authentication based on the probability of voice spoofing. Decision module 214 approves and / or rejects user authentication based on the user's voice input. Optionally, when decision module 214 approves user authentication, the learning engine can perform one or more operations based on the user's voice input. Optionally, when decision module 214 rejects user authentication, the learning engine causes the output module to remind the user.

[0054] Figure 3 This is a graphical representation 300 of the deception probability provided by one implementation of the present invention. Graphical representation 300 illustrates the relationship between the deception probability and distance, noise type, and noise level:

[0055] Probability of deception = f(distance, noise type, noise level).

[0056] Optionally, one or more features related to the Lombard effect in the time domain, spectral domain, and power spectral domain are retrieved from the user's speech input. Optionally, the one or more features include the speech fundamental frequency (F0, F1, F2), energy (EL, EM, EH) from the low-frequency to the mid-frequency or high-frequency bands, sound intensity (M), vowel duration (VD), spectral tilt (C), format center frequency (FFT×nBins), keyword duration (D1, D2…Dn, related to key phrases), etc. Optionally, the average estimate of one or more features is created for the user speaking under suitable conditions within a predefined period or according to a predefined number of recent conversations. Suitable conditions can be defined by noise type and signal-to-noise ratio (SNR) level. Processable noise includes mechanical noise (including noise from engines, air conditioners, vibrations, etc.), road noise (including noise from tires, wind, road vibrations, etc.), music noise, chat noise, and natural noise (including noise from wind, rain, forests, etc.). Distance can be calculated using the measured (FMi) and the estimated sample mean (FEi). Distance can be calculated using the following formula:

[0057]

[0058] The computational distance between the measured (FMi) and the estimated sample mean (FEi) can be used to estimate the probability of deception. Alternatively, advanced weighted distance (Wi) or alternative techniques can be used to improve accuracy. Alternative techniques can include machine learning techniques.

[0059] Figure 4 This invention provides a system architecture 400 for a device used for voice spoofing detection in noisy environments, according to one implementation. System architecture 400 includes a spoofing classification subsystem 402, a noise feature retrieval unit 404, and a training and adjustment subsystem 406. The spoofing classification subsystem 402 includes a Lombard feature retrieval unit 408 and an anomaly detector 410. The Lombard feature retrieval unit 408 receives user voice input and ambient noise. The user's voice input and ambient noise can be an input audio stream. Optionally, the Lombard feature retrieval unit 408 receives the input audio stream through an input module. The Lombard feature retrieval unit 408 is used to retrieve voice features from the input audio stream. Voice features can be any of the user's speech or commands.

[0060] A noise feature retrieval unit 404 is used to infer the type of noise in the input audio stream. The noise feature retrieval unit 404 can calculate noise parameters based on the noise type of the input audio stream. Optionally, the noise parameters include the signal-to-noise ratio (SNR). An anomaly detector 410 is used to compare the retrieved feature vector with the user's expected value in the noise context to determine the probability of deception. Optionally, the retrieved feature vector is speech features retrieved from the input audio stream. The anomaly detector 410 performs the comparison using feature templates. Optionally, the feature templates are stored in a Lombard template database. The Lombard template database can store Lombard features acquired and averaged during the training of the learning engine.

[0061] Optionally, the above-mentioned device further includes a data acquisition unit 412, which is used to acquire, for each user, features of the user's speech under normal conditions, noise parameters, actual values ​​of the user's speech Lombard effect-related features under noise parameters, and Lombard feature effect-related templates. Optionally, the above-mentioned device further includes a learning engine 414, which is used to: update the user's Lombard effect-related feature templates using the actual values ​​of the Lombard effect-related features in response to the decision module and / or learning trigger verifying the user's authentication.

[0062] The training and tuning subsystem 406 in system architecture 400 includes a data collector 412 and a learning engine 414. Optionally, the learning engine 414 is a user learning engine. The training and tuning subsystem 406 can receive new blocks through a new Application Programming Interface (API). The data collector 412 is used to create feature vectors for speech segments and manage the user's measurement set based on noise type and noise parameters in the input audio stream. Optionally, the creation of feature vectors can handle trends in the user's speech features. The data collector 412 can store the feature vectors as up-to-date data in a database. The learning engine 414 is used to refresh Lombard features stored in a Lombard template database for comparison by the anomaly detector 410. Optionally, the learning engine 414 can run periodically or be activated based on key performance indicators (KPIs) degradation, etc.

[0063] Optionally, system architecture 400 includes a voice verification system for authenticating users based on their voice input and the characteristics of their voice. The voice verification system can determine a user ID based on the user's voice input. Optionally, system architecture 400 includes deception classification, wherein the characteristics of the user's voice can be compared with expected Lombard characteristics. Optionally, system architecture 400 includes an additional system that runs a background learning and updating a Lombard template database that changes over time.

[0064] Figure 5A and Figure 5B This is a flowchart of a method for detecting speech spoofing in a noisy environment, provided by one implementation of the present invention. In step 502, a user's voice input is received. In step 504, an environmental noise measurement value corresponding to the user's voice input is received. In step 506, the user is authenticated based on the voice input. In step 508, noise parameters are determined based on the environmental noise measurement value. In step 510, the actual value of the Lombard effect-related feature of the user's voice is determined based on the user's voice input. In step 512, the probability of speech spoofing is determined by comparing the actual value of the Lombard effect-related feature with the Lombard effect-related feature template of the authenticated user corresponding to the determined noise parameters.

[0065] This method offers better protection against voice spoofing attacks in noisy environments. It also prevents replay, imitation, and speech synthesis attacks. The method can include additional independent authentication factors to authenticate users, thereby ensuring system security. It is relatively lightweight. Even in very noisy environments, it improves the accuracy of user voice input by increasing the probability of voice spoofing. Furthermore, it introduces a complementary technique that utilizes noise for voice spoofing detection, significantly enhancing voice anti-spoofing capabilities in noisy environments.

[0066] Optionally, the Lombard effect-related features include changes in the user's speech characteristics in the time domain, spectral domain, and power spectral domain under normal conditions and in the noisy environment, including changes in speech fundamental frequency, energy distribution between high, mid, and low frequency bands, sound intensity, vowel duration, spectral tilt, format center frequency, keyword duration, and lung volume.

[0067] Optionally, the noise parameters include noise type and noise level.

[0068] Optionally, the Lombard effect-related feature template includes preset values ​​of the Lombard effect-related features of the user's speech under different noise parameters.

[0069] Optionally, the preset value of the Lombard effect-related feature includes the average value of the Lombard effect-related feature collected for the user's speech under different noise parameters during the training period.

[0070] Optionally, based on the Lombard effect-related feature template corresponding to the determined noise parameters of the authenticated user, the voice spoofing probability is determined as a function of the difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature.

[0071] Optionally, the Lombard effect-related feature template further includes weights for the Lombard effect-related features, and the speech deception probability is determined as a function of the weighted difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, wherein the weighted difference is obtained through the weights.

[0072] Optionally, the method further includes: verifying the user's authentication based on the voice spoofing probability.

[0073] Optionally, the method further includes: in response to verifying the user's authentication, updating the user's Lombard effect-related feature template using the actual value of the Lombard effect-related feature.

[0074] It should be understood that the component arrangements shown in the described figures are exemplary, and other arrangements may exist. It should also be understood that the various system components (and elements) defined by the claims, described below, and shown in the various block diagrams represent components in some systems configured according to the subject matter disclosed herein. For example, one or more of these system components (and elements) may be implemented wholly or partially by at least some of the components illustrated in the arrangements shown in the described figures.

[0075] In addition, while at least one of these components is implemented at least partially as an electronic hardware component and thus constitutes a machine, the other components may be implemented in software, which, when included in the execution environment, constitutes a machine, hardware, or a combination of software and hardware.

[0076] While the invention and its advantages have been described in detail, it should be understood that various changes, substitutions and alterations may be made herein without departing from the scope and spirit of the invention as defined by the appended claims.

Claims

1. A method for detecting speech spoofing in a noisy environment, characterized in that, The method includes: Receive user voice input; Receive environmental noise measurement values ​​corresponding to the user's voice input; The user is authenticated based on the voice input; The noise parameters are determined based on the environmental noise measurements. The actual values ​​of the Lombard effect-related features of the user's speech are determined based on the user's speech input. The Lombard effect-related features include the changes in the characteristics of the user's speech in the time domain, frequency domain, and power spectrum domain under normal conditions and in the noisy environment, including changes in the fundamental frequency of the speech, the energy distribution between the high, mid, and low frequency bands, sound intensity, vowel duration, spectral tilt, format center frequency, keyword duration, and lung volume. The probability of speech spoofing is determined by comparing the actual values ​​of the Lombard effect-related features with the Lombard effect-related feature templates of the authenticated user corresponding to the determined noise parameters.

2. The method according to claim 1, characterized in that, The noise parameters include noise type and noise level.

3. The method according to claim 1, characterized in that, The Lombard effect-related feature template includes preset values ​​of the Lombard effect-related features of the user's speech under different noise parameters.

4. The method according to claim 3, characterized in that, The preset value of the Lombard effect-related feature includes the average value of the Lombard effect-related feature collected during the training period for the user's speech under different noise parameters.

5. The method according to claim 3, characterized in that, Based on the Lombard effect-related feature template corresponding to the determined noise parameters of the authenticated user, the voice spoofing probability is determined as a function of the difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature.

6. The method according to claim 5, characterized in that, The Lombard effect-related feature template also includes weights for the Lombard effect-related features. The probability of speech deception is determined as a function of the weighted difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, wherein the weighted difference is obtained through the weights.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: verifying the user's authentication based on the voice spoofing probability.

8. The method according to claim 7, characterized in that, The method further includes: in response to verifying the user's authentication, updating the user's Lombard effect-related feature template using the actual value of the Lombard effect-related feature.

9. An apparatus (102, 202) for detecting voice spoofing in noisy environments, characterized in that, The apparatus (102, 202) includes: Input modules (104, 204) are used to receive user voice input and environmental noise measurement values ​​corresponding to the user's voice input; A voice verification module (106) is used to authenticate the user based on the voice input; A noise detector (108) is used to determine noise parameters based on the environmental noise measurement values; A Lombard feature retrieval unit (110) is used to determine the actual values ​​of Lombard effect-related features based on the user's speech input. The Lombard feature retrieval unit (110) is used to determine the actual values ​​of the Lombard effect-related features by estimating the changes in the features of the user's speech in the time domain, spectral domain, and power spectral domain under normal conditions and in a noisy environment. The changes include changes in the fundamental frequency of the speech, the energy distribution between the high, mid, and low frequency bands, the sound intensity, the vowel duration, the spectral tilt, the format center frequency, the keyword duration, and the lung volume. A spoofing detector (112) is used to determine the probability of speech spoofing by comparing the actual value of the Lombard effect-related feature with a Lombard effect-related feature template corresponding to a determined noise parameter of an authenticated user.

10. The apparatus (102, 202) according to claim 9, characterized in that, The noise retrieval device (108) is used to determine the noise type and noise level.

11. The apparatus (102, 202) according to claim 9, characterized in that, The Lombard effect-related feature template includes preset values ​​of the Lombard effect-related features of the user's speech under different noise parameters.

12. The apparatus (102, 202) according to claim 11, characterized in that, The preset value of the Lombard effect-related feature includes the average value of the Lombard effect-related feature collected during the training period for the user's speech under different noise parameters.

13. The apparatus (102, 202) according to claim 11, characterized in that, The spoofing detector (112) is used to determine the speech spoofing probability as a function of the difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, based on the Lombard effect-related feature template corresponding to the determined noise parameters of the authenticated user.

14. The apparatus (102, 202) according to claim 13, characterized in that, The Lombard effect-related feature template also includes weights of the Lombard effect-related features. The deception detector (112) is used to determine the speech deception probability as a function of the weighted difference between the actual value of the Lombard effect-related feature and the preset value of the Lombard effect-related feature, wherein the weighted difference is obtained through the weights.

15. The apparatus (102, 202) according to any one of claims 9 to 14, characterized in that, The device (102, 202) further includes a decision module (214) for verifying the user's authentication based on the voice spoofing probability.

16. The apparatus (102, 202) according to claim 15, characterized in that, The device (102, 202) also includes a data acquisition unit, which is used to collect for each user the features of the user's speech under normal conditions, noise parameters, the actual values ​​of the Lombard effect-related features of the user's speech under the noise parameters, and the Lombard effect-related template.

17. The apparatus (102, 202) according to claim 16, characterized in that, The apparatus (102, 202) further includes a learning engine (414) for: updating the user's Lombard effect-related feature template using the actual values ​​of the Lombard effect-related features in response to the decision module and / or learning trigger verifying the user's authentication.

Images