Obtaining impulse response of room

Abstract

It is proposed to process sound signals acquired by an array of microphones and generated by a sound source to acoustically characterize a space (ESP) comprising the array (MIC) and the sound source and delimited by walls (PAR). A time-frequency transform is applied to the acquired signal and a generalized velocity vector is expressed in the frequency domain based on the acquired signal. Specifically, the generalized velocity vector is represented in the time domain v (t) in the form of a series of peaks comprising at least one peak associated with reflections from the wall and time abscissas associated with the delay TAU1, and the expression of the generalized velocity vector in the time domain is modeled by autoregressive moving average ARMA, the autoregressive moving average (ARMA) is defined by an autoregressive filter (AR) and a moving average (MA). Thus, it is provided that the acquired sound signal is processed by applying an autoregressive filter (AR) to obtain an impulse response characterizing the space (ESP) and obtained from the moving average (MA).

Description

Technical Field

[0001] This specification relates to the field of sound data processing. More specifically, this specification relates to obtaining the impulse response of a room (enclosed space) based on an impulse response known as the "generalized relative impulse response". Background Technology

[0002] Knowledge of the acoustic and geometric properties of an environment can make it possible to obtain relevant results in audio signal processing, or to improve those results for a variety of use cases. Simultaneous audio processing, including locating and isolating sound sources in the environment, can be advantageous, especially under adverse conditions (e.g., in the presence of obstacles that prevent the propagation of sound in a straight line). There is a strong demand for such processing, particularly in applications such as spatial coding, augmented reality, robot navigation, and room representation.

[0003] When using sound-based methods to estimate acoustic environments, it is often necessary to leverage the capabilities of multiple microphones that encode spatial information. One particularly suitable representation of a 3D sound field is the Higher Order Ambisonics (HOA) format, hereinafter referred to as "ambisonics," which is based on decomposing sound pressure on a sphere into spherical harmonics. Ambisonic channels are coherent with each other, but they differ in directionality; that is, they have different sensitivities to excitations from different spatial directions. They can be recorded by specific devices (most commonly spherical microphone arrays or "SMAs") or artificially created. In a given environment, and for a given source and a given SMA device location, each HOA channel has a specific room impulse response (abbreviated as RIR). These RIR responses provide information about the environment in which sound propagates, particularly in the first part of the response (i.e., in the "first echo").

[0004] Even if a spatial fingerprint is embedded in the recorded audio, recovering that information is not straightforward. On one hand, the RIR is linked to the (unknown) source signal, and on the other hand, the recording may be contaminated with noise. For this reason, all inference methods base their analysis on a set of pre-recorded or estimated RIRs (not necessarily in HOA format). While such analysis can be inherently challenging (e.g., due to the problem of “marking” echoes, which aims to attribute each signal peak to reflections from a partition), this is a very large assumption that would limit its application to use cases where only RIRs are available. To circumvent this problem, blind inference methods can be based on analyzing a so-called “phase-aligned” spatial correlation matrix. However, the computational complexity of this approach appears excessive for real-time processing.

[0005] Alternatively, relative fingerprints, i.e., relative transfer functions (referred to as "ReTF" in the frequency domain) or relative impulse responses (referred to as "ReIR" in the time domain), can be considered to infer environmental properties. ReTF and ReIR model the relationship between each channel and a given reference signal, typically chosen as one of the channels. Theoretically, these representations are source-independent, but the cost of this approach is the inevitable loss of some information (specifically, the propagation time and absolute attenuation of the signal directly from the source to the microphone). Typically, ReIR responses are not causal, and their analysis is significantly more complex than that of RIR responses.

[0006] However, in the work corresponding to document WO-2022 / 106765, it has been demonstrated that using a reference signal (i.e., reference beamforming) as a linear combination of all channels is advantageous when extracting information from the ReIR of a stereo reverberation signal. More specifically, if beamforming sufficiently attenuates acoustic reflections compared to direct propagation, the corresponding ReIR (called the “generalized time-domain velocity vector” and hereinafter denoted as “GTVV”) allows for a rich and compact representation of information in the time domain. Under these conditions, the generalized ReIR is causal and relatively sparse, and therefore allows for estimation of the direction of acoustic reflections and their associated delays based on the peak value of the direction of arrival of the sound (denoted as DoA).

[0007] The GTVV vector (i.e., the ReIR in the stereo reverberation domain with beamforming) is more robust to adverse acoustic conditions than the "standard" ReIR, where the reference signal is typically a zero-order omnidirectional stereo reverberation channel. However, it is limited by the performance of the beamforming applied. For example, if signal-independent maximum directivity factor beamforming is used, its directivity is a quadratic function of a given HOA order. However, conventional SMA microphone equipment typically does not provide sufficiently high-order stereo reverberation formats: most commonly, they are only able to record first-order stereo reverberation signals (FOA). This is especially true for simple devices that only support FOA (e.g., handheld devices). Furthermore, frequency support for higher-order channels decreases progressively with HOA order as noise amplification at low frequencies and spatial aliasing at high frequencies begin to appear.

[0008] However, because beamforming cannot effectively suppress reflections, the favorable theoretical properties of the GTVV vector tend to decrease at low stereo reverberation orders. This problem is further exacerbated as the distance between the microphone and the source increases, because more reflections fall within the main beamforming lobe, and the advantage of direct sound relative to reflections decreases simultaneously. In practice, the GTVV fingerprint can be observed to be no longer causal, and the estimated direction becomes less accurate.

[0009] Furthermore, even when the GTVV representation remains valid, extracting direction and delay via peak identification / selection is not necessarily direct. An easy-to-use GTVV vector can be considered as a multi-channel RIR response (no delay, centered) implied by a causal filter. As a result, at moments corresponding to integer multiples of their relative delay, the same reflections repeat infinitely as echoes, with alternating signs and decreasing amplitudes. Therefore, this series can interfere with each other, altering the information that can be derived from them, or even masking the presence of, for example, lower-amplitude reflections. Summary of the Invention

[0010] This instruction manual improves upon this situation.

[0011] To this end, a method is proposed for processing sound signals acquired by at least one microphone array and originating from at least one sound source, in order to acoustically characterize a space comprising the array and the source and defined by at least one wall.

[0012] in:

[0013] - Apply time-frequency transformation to the acquired signal.

[0014] - Based on the acquired signal, the generalized velocity vector V(f) is represented in the frequency domain. It is a complex number with real and imaginary parts, and the velocity vector represents a combination of the following terms:

[0015] *The first acoustic path, which lies directly between the source and the microphone array, is represented by the first vector U0, and

[0016] *At least a second acoustic path, which originates from reflections on the wall and is represented by a second vector U1.

[0017] The second path has a delay TAU1 at the microphone array relative to the direct path.

[0018] - Apply an inverse transformation from frequency to time to the generalized velocity vector so that the generalized velocity vector is represented in the time domain v(t) as a series of peaks, including at least one peak associated with the reflection on the wall and the time abscissa as a function of the delay TAU1.

[0019] Specifically, in this method, the expression of the generalized velocity vector in the time domain is modeled by an autoregressive moving average (ARMA) defined by an autoregressive filter (AR) and a moving average (MA), and the method then includes processing the acquired acoustic signal by applying the autoregressive filter (AR) to obtain an impulse response characterizing the space and derived from the moving average (MA).

[0020] With this arrangement, information stored in the representation of the generalized velocity vector (referred to below as “GTVV”) in the time domain can be extracted more robustly, as it is more general for any acoustic situation to obtain an impulse response characterizing a space with at least one wall (such as a room, and thus corresponding to an impulse response of the RIR (“room impulse response”) type). More specifically, as described below in an exemplary embodiment, this impulse response can be described as “reduced” (and denoted as “RdRIR”, representing “reduced room impulse response”), because the time expression of the generalized velocity vector from which this impulse response is derived only provides the reflection delay (not the absolute delay) relative to the reception delay of the direct acoustic path from the source at the microphone. Similarly, the amplitude of the reflection is relative to the amplitude of the received direct sound (not reflected by the walls) at the microphone. However, simply by treating the time expression of the generalized velocity vector as an ARMA model, such an impulse response (even relative) already allows for efficient characterization of the acoustic space under consideration.

[0021] Therefore, the reduced impulse response RdRIR comes from the ARMA model and is thus distinguished from the previously introduced relative impulse response ReIR, which is obtained directly from the expression of the generalized velocity vector.

[0022] In one embodiment, the acquired signal is applied to the stereo reverberation channel, and the autoregressive filter (AR) is common to all channels.

[0023] This implementation scheme in stereo reverberation representation has the advantage of not requiring a very high stereo reverberation order (a first-order or "FOA" representing "first-order stereo reverberation" is sufficient to obtain a satisfactory impulse response).

[0024] In the embodiment where the space is defined by multiple walls, the time-domain expression of the generalized velocity vector includes a series of peaks, including a peak associated with the direct path (or "DoA," representing "direction of arrival"), followed by peaks each associated with at least one reflection on wall n. The method then includes:

[0025] - Optimize the autoregressive filter to model the series of peaks in the form of a multivariate autoregressive moving average.

[0026] Therefore, the time representation of the generalized velocity vector is well-suited for modeling via multivariate ARMA averaging.

[0027] In particular, in such an embodiment, the method may include:

[0028] - Based on the expression of the generalized velocity vector in the time domain v(t) in the form of the aforementioned series of peaks, the autoregressive filter is optimized by utilizing the causal properties of the impulse response.

[0029] In fact, the generalized velocity vector can be represented in the time domain as follows:

[0030]

[0031] Where h(t) and a(t) represent the causal filters of the moving average part (MA) and the regression part (AR) of the ARMA model, respectively, and through... Associated, representing beamforming w received by the microphone array in the direction of arrival of the sound from the sound source, and wherein:

[0032] y n g n ∈]0,1[and τ n >0 indicates the parameter of the nth plane wave reflected by the nth wall in space.

[0033] y n It is the vector encoding the directional propagation of the nth wave.

[0034] g n It is the relative attenuation of the nth wave, and

[0035] τ n It is the delay of the nth wave relative to the direct path.

[0036] y0 is the propagation vector specific to the direct path.

[0037] It is the beamforming response of the nth wavefront, where β0 = 1.

[0038] In this implementation, the autoregressive component a(t) can then be estimated by minimizing the following equation:

[0039]

[0040] With the constraint a(0) = 1, where v l (t) is a channel of the generalized velocity vector, represented by a multivariate ARMA model, where the autoregressive component a(t) is common to all channels of the generalized velocity vector.

[0041] In such an embodiment, estimating a(t) by minimizing ε is equivalent to solving a linear predictive system, which is advantageously overdetermined.

[0042] Using the notation presented above, the impulse response can be given by the moving average h(t) such that: h(t) = (v*a)(t).

[0043] Since the time representation of the generalized velocity vector can have positive and negative amplitudes (e.g., Figure 2 As shown in the example, the amplitude sign can be corrected by applying a correction to the moving average h(t) to obtain the general expression of the impulse response in its positive form.

[0044] Furthermore, in one embodiment, the impulse response is selected to have a finite duration. This duration can be specifically chosen to avoid taking into account the diffuse reverberation field (typically, it will be within the range of the diffuse reverberation field). Figure 2 (Multiple higher-order reflections appearing on the far right of the image), and therefore only early reflections on the walls of the space under consideration are dealt with.

[0045] In such an embodiment, the property that the aforementioned impulse response has a finite duration can be used to set a limit on the maximum filter length of the autoregressive (AR) portion and the moving average (MA) portion.

[0046] In another embodiment, the sliding average portion MA can be centered on a delay that corresponds to the time it takes to receive sound from the source at the microphone.

[0047] On the other hand, a computer program comprising instructions that, when executed by processing circuitry, implement the aforementioned method is proposed. Furthermore, a non-transitory computer-readable storage medium on which such a program is stored is also proposed.

[0048] On the other hand, it also relates to a device including processing circuitry, which includes an interface for receiving sound signal data acquired by a microphone array and is configured to implement the above-described method. Attached Figure Description

[0049] Other features, details, and advantages will become apparent from reading the following detailed description and analyzing the accompanying drawings, in which:

[0050] Figure 1 An example of a series of steps for a method of the above type is shown.

[0051] Figure 2 An example of the time representation of the generalized velocity vector is shown.

[0052] Figure 3 The diagram shows real-world examples of the generalized time-domain velocity vector (in the middle), the corresponding ARMA representation (on the right), and the true impulse response (on the left) under different conditions. The ARMA model is more faithful to the true impulse response.

[0053] Figure 4 , Figure 5 and Figure 6 The diagram illustrates the performance of the results obtained by implementing the above method compared to other treatments (or no treatment) for different corresponding durations of the sound reverberation period.

[0054] Figure 7 It shows something similar to Figures 4 to 6 The evaluation of the error in the direction of arrival (DoA) of sound under experimental conditions more clearly demonstrates the performance of the above method, especially for reflection.

[0055] Figure 8 The apparatus used to implement this method is shown schematically. Detailed Implementation

[0056] We first refer to Figure 1 It illustrates the steps of a method of the above type according to an exemplary embodiment.

[0057] We first describe Figure 1 The main principle of the steps in the process.

[0058] Stereo reverberation signals of any order are recorded, for example, by a device such as an SMA microphone array (or generated by analog or other means). These multichannel signals are then used to estimate the generalized time-domain velocity vector (GTVV). This estimation is typically performed in practice by computing the inverse Fourier transform of the corresponding relative transfer function (ReTF) in the frequency domain, as specifically described in document WO-2022 / 106765. Therefore, the stereo reverberation signal is usually pre-transformed into a time-frequency representation (e.g., via STFT, standing for "Short-Time Fourier Transform"), and a robust estimator is used to obtain the ReTF.

[0059] The time shape of the GTVV vector Figure 2 Presented for illustrative purposes, and showing:

[0060] - The zero-delay peak associated with the direct acoustic path, which is related to the main DoA (“direction of arrival”) of the sound, and

[0061] - A peak associated with higher latency and with reflections on the partition.

[0062] The temporal fingerprint of the GTVV vector can be considered as an implementation of a multivariate autoregressive moving average or "ARMA" process, where the AR autoregressive filter (in the denominator) is common to all channels.

[0063] Therefore, once the GTVV vector is obtained, computation is performed by estimating the parameters of the corresponding ARMA model. By leveraging the fact that the RIR response is causal (as in the case of AR and MA filters in ARMA models), the AR filter is first estimated from the time series given by the GTVV vector, such as... Figure 2 As an example, the above is taken from the aforementioned document WO-2022 / 106765.

[0064] The maximum length of the AR and MA filters can be further limited because it is assumed that the first part of the RIR has a finite duration; in fact, such an AR filter can be efficiently computed by estimating a linear prediction model applied to the appropriate part of the GTVV fingerprint.

[0065] Once the AR filter is available, the MA filter can be estimated simply by convolving the GTVV vector with the AR filter (an efficient estimation, such as the Prony estimation) or by least-squares estimation (such as the Shanks estimation method); more advanced inference processes can also be considered here, such as by applying some structure between the corresponding inputs of the MA filter.

[0066] Ideally, an MA filter should approximate the normalized RIR response, where the main peak is centered at zero delay in the representation (absolute gain and delay remain unchanged due to information loss in ReIR, as previously stated). Due to its similarity to RIR, a sequence of such MA filters is called RdRIR, representing a “reduced room impulse response.” In practice, the RIR response is considered to have a positive amplitude and be a continuous function in time. When represented by a discrete (multichannel) time series, the RIR response is pre-filtered by an anti-aliasing filter, which typically presents an impulse response containing both positive and negative amplitudes. Since the same filter is applied to all HOA channels, the sign of the estimated zero-order RdRIR response can be observed, and (if negative) the signs of all channels for a given time sample can be inverted.

[0067] After correcting the sign of the RdRIR representation, wavefront inference can be performed. A non-limiting way to do this is to select the peak of a series of amplitude peaks in the RdRIR response at different times.

[0068] The following section describes in more detail the application of ARMA modeling to the GTVV (Generalized ReIR) vector in the stereo reverberation domain.

[0069] The proposed principle can be applied to standard (non-stereo reverberant) ReIR responses, for example, to estimate the time difference of arrival (or TdoA) when considering a pair of microphones recording signals from the same source.

[0070] The following mathematical description covers the definition of GTVV vectors, wavefront inference when GTVV vectors are theoretically valid (i.e., when the convergence conditions explained below are met), and the derivation of the ARMA-based "GTVV preconditioning" method presented above.

[0071] In the following text, we will denote the vector (denoted as "SH") of the spherical harmonic expansion coefficients as: It is cascaded at frequency f (corresponding to the "HOA channel") up to order L. Assume the recorded signal is from a far-field source in an indoor environment (typically a partitioned room) at azimuth angle θ0, elevation angle φ0, and distance d0 from the SMA microphone array. Given a (approximate) broadband beamforming w (or "beamforming" hereinafter) pointing to DoA, as specifically described in WO-2022 / 106765, the generalized frequency domain velocity vector (GFVV) is defined as follows:

[0072]

[0073] Where y n g n ∈]0,1[and τ n >0 represents the parameters of the nth plane wave reflected by the room's partitions, where:

[0074] y n In the direction (θ) n , φ n The expanded vector SH on )

[0075] g n It is its relative decay and

[0076] τ n It is its delay (compared to the direct propagation component).

[0077] Next, y0 is the vector SH of the plane wave on the direction DoA given by (θ0, φ0), and It is the response of the beamformer to the nth wavefront (where β0 = 1).

[0078] This approximation is due to the simplifying assumptions included on the right-hand side of the above equation: the plane wave decomposition has been given based on the main acoustic reflections, and it is assumed that beamforming and relative attenuation are independent of frequency.

[0079] Inverse Fourier Transform of Each Channel of the GFVV Vector Provide the corresponding GTVV time item:

[0080] In practice, the processing is performed in the STFT (Short Time Fourier Transform) domain, and the duration of the GTVV is determined by the selected window. The window of length T is centered at t=0 relative to the GTVV, i.e., t∈[-T / 2,T / 2].

[0081] Conditions for convergence of (geometric) Taylor series Under this, GTVV allows expressions of the following forms:

[0082]

[0083] η(t) accumulates “cross terms” (which involve mutual interference between different wavefronts).

[0084] The above expression "Equation 1" allows the wavefront of the direct sound y0 to be estimated immediately by evaluating v (t = 0), while the rest involves the summation of an infinite series corresponding to the reflected wavefront n ∈ [1, N].

[0085] However, due to g n |β n |<1, each infinite series has a peak time position t=τ n And the magnitude of the reduction:

[0086]

[0087] When beamforming has high selectivity, its response is β. n →0, therefore v(τ) n )≈g n y n If this is not the case, then it can be achieved by adjusting the observed vector v(τ). n To improve y, we need to perform "debiasing". n The estimate.

[0088] Given an estimate of y0 and a set of vectors SH {y n′}(corresponding to the direction set {(θ) n′ , φ n′ )} n′ ), and know g n It is strictly positive, which can be achieved by finding v(τ) in equation 2. n Retrieve y by maximizing the relevance of the elements. n It is written as:

[0089]

[0090] Alternatively, we can employ nonlinear optimization and solve Equation 2 in parametric form, where y n′ Transformed into a direction variable (θ) n′, φ n′ The function of ).

[0091] Convergence condition: This suggests that beamforming significantly attenuates reflections, depending on the type of beamforming applied, but also, of course, on the acoustic environment and the order of the HOA.

[0092] For computational reasons, it is convenient to use simple beamforming, such as the maximum directional factor beamforming w given by w:=y0 (in N3D stereo reverberation coding, having enough knowledge to weight the signal acquired by a stereo reverberation microphone (with multiple piezoelectric capsules to collect multiple sound signals) to switch from one coding type to another).

[0093] However, due to the width of its main lobe, this beamforming is too loose under low-order stereo reverberation (e.g., multiple FOAs), so the expression in Equation 1 may no longer be valid.

[0094] However, the GTVV vector can still be written in the following form:

[0095]

[0096] Where h(t) and a(t) are both causal filters, which pass through Related.

[0097] This expression reveals a specific structure (v per GTVV channel). l (t) can be viewed as an implementation of a multivariate ARMA model, where the autoregressive component a(t) is common to all channels.

[0098] Therefore, the series h(t) of the MA portion, or the RdRIR response (reduced room impulse response), allows for expressions of the following type:

[0099] h(t) = (v*a)(t). Equation 4

[0100] Since h(t<0)=0, we can estimate a(t) by minimizing the following equation.

[0101]

[0102] It has the constraint a(0) = 1.

[0103] This is advantageously an overdetermined problem: the length of the filter a(t) is T / 2, while the number of data points is (L+1). 2 T / 2 (the non-causal part of the GTVV vector representing v(t)). As the order L of the HOA increases, the estimation should become more accurate because more data becomes available for regression.

[0104] The cost function can be extended to incorporate the weight α. l and the last part of the wavefront sequence (t>t) max It is assumed to be a low-amplitude signal similar to noise:

[0105]

[0106] Since the two filters are derived from linear expressions Related, therefore, by applying the condition h(t>t) max Since ) = 0, the support for filter a(t) is therefore implicitly shortened from T / 2 to t. max .

[0107] In principle, by further modifying the original cost function, it is possible to incorporate more structure into h(t) (or a(t)). One such example could be using a norm that favors the sparsity of the group to support h(0≤t≤t). max Modeling is then performed. However, solving such optimization problems typically requires additional computational resources. Therefore, least squares minimization is presented here as an example.

[0108] Taking the partial derivative of ε with respect to the elements of the AR filter and expressing it as a(t″), and setting the result to zero, we obtain:

[0109]

[0110] in

[0111]

[0112] The two autocorrelation functions defined above can be efficiently computed using the Fast Fourier Transform. Their total (weighted) sum can be written as:

[0113]

[0114] Since α(0) = 1, estimating the residual coefficients is equivalent to the classic linear prediction problem:

[0115]

[0116] This can be solved in various ways.

[0117] For example, in order to use a fast solver based on Toplitz, the original cost function can be slightly modified, and used as an alternative function to minimize the following type:

[0118]

[0119] in:

[0120]

[0121] and

[0122]

[0123] Once a(t) has been computed, the non-zero segment (RdRIR) of h(t) can be retrieved by evaluating (v*a) (t≥0).

[0124] This implementation is computationally very efficient. However, more refined methods can be applied, such as estimating RdRIR in the least-squares sense (the so-called Shanks method), or even performing alternating optimization to improve AR and RdRIR (known as the Steiglitz-McBride algorithm). These methods require estimating the inverse AR filter, which is typically approximated by an FIR filter that is optimal in the least-squares sense.

[0125] The characteristics are represented in matrix form. Given, where H represents the sequence of GTVV vectors from Equation 1 or the estimated RdRIR sequence h(t) from Equation 3.

[0126] exist Figure 3 Examples of such sequences are given, for recordings of FOAs collected from speech sources from SMA devices, and for real multichannel RIR responses (shifted so that their main peak is at t=0). Clearly, the proposed RdRIR then approximates the RIR structure more closely than the fingerprint of the GTVV vector.

[0127] DoA is evaluated based on the vector in matrix H corresponding to zero delay, while the remaining directions are obtained by selecting the amplitude peaks of its column vectors. The index of the selected peaks reveals the relative delay of a given direction with respect to the direct path of the sound.

[0128] Next, the angular error of the direction associated with the ten largest peaks in the corresponding sequence can be quantified.

[0129] Three methods are considered below:

[0130] - None (no post-processing)

[0131] - Bias correction (using the bias correction in Equation 3), and

[0132] -In the sense of the above method, it has ARMA correction (RdRIR).

[0133] Execute for different stereo reverberation orders L=1, L=2, or L=3.

[0134] More specifically, evaluations were performed for HOA orders L=1, L=2, and L=3, with SNRs of 0dB, 10dB, 20dB, and “infinite (Inf)” dB (i.e., virtually no noise). Each result is a median estimate of 10 replicates for a given simulation configuration (i.e., for a given reverberation time and level of additive white Gaussian noise). The experimental simulation size was 5×4×3m. 3 The rectangular room contains microphone arrays and voice sources that are randomly positioned, but the distance between them is between 0.5m and 6m.

[0135] The experimental conditions for the three reverberation periods (RT60 = 200ms, RT60 = 400ms and RT60 = 600ms) were respectively... Figure 4 , Figure 5 and Figure 6 The results clearly show that RdRIR provides the most accurate estimate, and the performance of all methods increases with the HOA order and deteriorates with increasing reverberation time and noise level. However, it is noteworthy that RdRIR often outperforms the other methods even when its HOA order is lower than that of the other two methods.

[0136] Figure 7 The evaluation of DoA error alone is shown for RT60 = 400 ms under similar experimental conditions. Although the RdRIR estimate again shows the lowest angular error, the difference is not significant for all SNR levels. This indicates that the main improvement of ARMA post-processing lies in better prediction of the wavefront, especially the reflected wavefront.

[0137] Figure 8 An example of a device for implementing the above method is shown, and it typically includes:

[0138] - Interface INT, used to receive signals from a microphone MIC, such as a stereo reverberation microphone (e.g., with several piezoelectric capsules), the microphone MIC being arranged in a space ESP including at least one wall PAR.

[0139] - The processor PROC, connected to the interface INT, processes, for example, received signals represented in stereo reverberation, to express the generalized velocity vector in time as a function of these signals, and derives the ARMA model to transmit the impulse response RdRIR of the spatial ESP.

[0140] - Memory MEM stores the instruction data of a computer program as defined in this specification, and is accessible by processor PROC to read the data and execute the methods described above.

[0141] Obtaining the impulse response of a spatial ESP makes it possible to specifically quantify the acoustic and geometric properties of that space (e.g., simultaneously locating and separating sound sources within the spatial ESP, or others). Knowledge of the acoustic and geometric properties of this environmental ESP can enable the acquisition or improvement of relevant results in various applications of processing audio signals for spatial coding, augmented reality, robot navigation, room representation, etc., as mentioned above. Using an ARMA model to obtain the room's impulse response is readily achievable (especially for the desired low-order stereo reverberation) and yields satisfactory results, such as... Figures 4 to 7 As shown.

Claims

1. A method for processing sound signals acquired by at least one microphone array and originating from at least one sound source, so as to acoustically characterize a space comprising said array and said source and defined by at least one wall. in: - Apply time-frequency transformation to the acquired signal. - Based on the acquired signal, the generalized velocity vector V(f) is represented in the frequency domain. It is a complex number with real and imaginary parts, and the velocity vector represents a combination of the following terms: *The first acoustic path, which is directly between the source and the microphone array, is represented by the first vector U0, and *At least a second acoustic path, which originates from reflections on the wall and is represented by a second vector U1. The second path has a delay TAU1 at the microphone array relative to the direct path. - Apply an inverse frequency-to-time transform to the generalized velocity vector so that it is represented in the time domain v(t) as a series of peaks, including at least one peak associated with the reflection on the wall and the time abscissa as a function of delay TAU1. The time-domain expression of the generalized velocity vector is modeled by an autoregressive moving average (ARMA), which is defined by an autoregressive filter (AR) and a moving average (MA). The method involves processing the acquired sound signal by applying an autoregressive filter (AR) to obtain an impulse response characterizing the space and derived from a moving average (MA).

2. The method according to claim 1, wherein, The acquired signal is applied to the stereo reverberation channel, and the autoregressive filter (AR) is common to all channels.

3. The method according to any one of the preceding claims, wherein, For a space defined by multiple walls, the time-domain expression of the generalized velocity vector includes a series of peaks, including a peak associated with the direct path (DoA), followed by a peak each associated with at least one reflection on wall n. The method includes: - Optimize the autoregressive filter to model the series of peaks in the form of a multivariate autoregressive moving average.

4. The method according to claim 3, comprising: - The autoregressive filter is optimized by utilizing the causal properties of the impulse response, based on the expression of the generalized velocity vector in the time domain v(t) in the form of the series of peaks.

5. The method according to claim 4, wherein, The generalized velocity vector is represented in the time domain in the following form: , in and These are causal filters representing the moving average (MA) and regression (AR) components of the ARMA model, respectively, and are obtained through... Associated, representing beamforming w received by the microphone array in the direction of arrival of the sound from the sound source, and wherein: , and Let represent the parameters of the nth plane wave reflected by the nth wall in space. It is the vector encoding the directional propagation of the nth wave. It is the relative attenuation of the nth wave, and It is the delay of the nth wave relative to the direct path. It is a propagation vector specific to the direct path. It is the beamforming response of the nth wavefront, where .

6. The method according to claim 5, wherein, The autoregressive component is estimated by minimizing the following expression: , With constraints ,in It is the channel of the generalized velocity vector, represented by a multivariate ARMA model, and the autoregressive part All channels of the generalized velocity vector are common.

7. The method according to claim 6, wherein, By minimizing To estimate This is equivalent to solving an overdetermined linear prediction system.

8. The method according to any one of claims 5 to 7, wherein, The impulse response is determined by the moving average. Given that: 。 9. The method according to claim 8, wherein, For the moving average Amplitude sign correction is applied to obtain a positive impulse response.

10. The method according to any one of the preceding claims, wherein, The impulse response has a finite duration.

11. The method according to claim 10, wherein, Limits are set on the maximum filter length for the autoregressive component AR and the moving average component MA.

12. The method according to any one of the preceding claims, wherein, The sliding average portion MA is centered on a delay, which corresponds to the reception time of sound originating from the source at the microphone.

13. A computer program comprising instructions which, when executed by a processing circuit, are used to implement the method according to any one of the preceding claims.

14. An apparatus comprising processing circuitry including an interface for receiving audio signal data acquired via a network of microphones, and configured to implement the method according to any one of claims 1 to 12.

Images