The present invention relates generally to audio recording, and more specifically to the mixing, recording and playback of audio signals for reproducing real or virtual three-dimensional sound scenes at the eardrums of a listener using loudspeakers or headphones.
A well-known technique for artificially positioning a sound in a multi-channel loudspeaker playback system consists of weighting an audio signal by a set of amplifiers feeding each loudspeaker individually. This method, described e.g. in [Chowning71], is often referred to as “discrete amplitude panning” when only the loudspeakers closest to the target direction are assigned non-zero weights, as illustrated by the graph of panning functions in
W(σ,φ)=1
X(σ,φ)=cos(φ)cos(σ)
Y(σ,φ)=cos(φ)sin(σ)
Z(σ,φ)=sin(φ)
where σ and φ denote respectively the azimuth and elevation angles of the sound source with respect to the listener, expressed in radians. An advantage of this technique over the discrete panning method is that B Format encoding does not require knowledge of the loudspeaker layout, which is taken into account in the design of the decoder. A second advantage is that a real-world B-Format recording can be produced with practical microphone technology, known as the ‘Soundfield Microphone’ [Farrah79]. As illustrated in
3-D audio reproduction techniques which specifically aim at reproducing the acoustic pressure at the two ears of a listener are usually termed binaural techniques. This approach is illustrated in
Conventional binaural techniques can provide a more convincing 3-D audio reproduction, over headphones or loudspeakers, than the previously described techniques. However, they are not without their own drawbacks and difficulties.
[Travis96] describes a method for reducing the computational cost of the binaural synthesis and addresses the interpolation and dynamic issues. This method consists of combining a panning technique designed for N-channel loudspeaker playback and a set of N static binaural synthesis filter pairs to simulate N fixed directions (or “virtual loudspeakers”) for playback over headphones. This technique leads to the topology of
[Lowe95] describes a variation of the topology of
There remains a need for a computationally efficient technique for high-fidelity 3-D audio encoding and mixing of multiple audio signals. It is desirable to provide an encoding technique that produces a non listener-specific format. There is a need for a practical recording technique and suitably designed decoders to provide faithful reproduction of the pressure signals at the ears of a listener over headphones or two-channel and multi-channel loudspeaker playback systems.
A method for positioning an audio signal includes selecting a set of spatial functions and providing a set of amplifiers. The gains of the amplifiers being dependent on scaling factors associated with the spatial functions. An audio signal is received and a direction for the audio signal is determined. The scaling factors are adjusted depending on the direction. The amplifiers are applied to the audio signal to produce first encoded signals. The audio signal is then delayed. The second filters are then applied to the delayed signal to produce second encoded signals. The resulting encoded signals contain directional information. In one embodiment of the invention, the spatial functions are the spherical harmonic functions. The spherical harmonics may include zero-order and first-order harmonics and higher order harmonics. In another embodiment, the spatial functions include discrete panning functions.
Further in accordance with the method of the invention, a decoding of the directionally encoded audio includes providing a set of filters. The filters are defined based on the selected spatial functions.
An audio recording apparatus includes first and second multiplier circuits having adjustable gains. A source of an audio signal is provided, the audio signal having a time-varying direction associated therewith. The gains are adjusted based on the direction for the audio. A delay element inserts a delay into the audio signal. The audio and delayed audio are processed by the multiplier circuits, thereby creating directionally encoded signals. In one embodiment, an audio recording system comprises a pair of soundfield microphones for recording an audio source. The soundfield microphones are spaced apart at the positions of the ears of a notional listener.
According to the invention, a method for decoding includes deriving a set of spectral functions from preselected spatial functions. The resulting spectral functions are the basis for digital filters which comprise the decoder.
According to the invention, a decoder is provided comprising digital filters. The filters are defined based on the spatial functions selected for the encoding of the audio signal. The filters are arranged to produce output signals suitable for feeding into loudspeakers.
The present invention provides an efficient method for 3-D audio encoding and playback of multiple sound sources based on the linear decomposition of HRTF using spatial panning functions and spectral functions, which
The use of predetermined panning functions offers the following advantages over methods of the prior art which use principal components analysis or singular value decomposition to determine panning functions and spectral functions:
Two particularly advantageous choices for the panning functions are detailed, offering additional benefits:
Given a set of N spatial panning functions {gi(σ , φ), i=0, 1, . . . N−1} the procedure for modeling HRTF according to the present invention is as follows. This procedure is associated to the topologies described in
In practice, the measured HRTFs are obtained in the digital domain. Each HRTF is represented as a complex frequency response sampled at a given number of frequencies over a limited frequency range, or, equivalently, as a temporal impulse response sampled at a given sample rate. The HRTF set {L(σp, φp, f)} or {R(σp, φp, f)} is represented, in the above decomposition, as a complex function of frequency in which every sample is a function of the spatial variables σ and φ, and this function is represented as a weighted combination of the spatial functions gi(σ, φ). As a result, a sampled complex function of frequency is associated to each spatial function gi(σ, φ), which defines the sampled frequency response of the corresponding filter L1(f) or Ri(f). It is noted that, due to the linearity of the Fourier transform, an equivalent decomposition would be obtained if the frequency variable f were replaced by the time variable in order to reconstruct the time-domain representation of the HRTF.
The equalization and the symmetrization of the HRTF sets L(σp, φp, f) and R(σp, φp, f), are not necessary to carrying out the invention. However, performing these operations eliminates some of the artifacts associated to the HRTF measurement method. Thus, it may be preferable to perform these operations for their practical advantages.
Step 2 is optional and is associated to the binaural synthesis topologies described in
ITD(σ,φ)=tR(σ,φ)−tL(σ,φ).
It is noted that the above procedure differs from the methods of the prior art. Conventional analytical techniques, such as PCA and SVD, simultaneously produce the spectral functions and the spatial functions which minimize the least-squares error between the original HRTFs and the reconstructed HRTFs for a given number of channels N. In the elaboration of the present invention, it is recognized in particular, that these earlier methods suffer from the following drawbacks:
In comparison, the technique in accordance with the present invention permits a priori selection of the spatial functions, from which the spectral functions are derived. As will be apparent from the following description, several benefits of the present invention will result from the possibility of choosing the panning functions a priori and from using a variety of techniques to derive the associated reconstruction filters.
An immediate advantage of the invention is that the encoding format in which sounds are mixed in
Generally, it is possible to make a selection of spatial panning functions and tune the reconstruction filters to achieve practical advantages such as:
Two particular choices of spatial panning functions will be detailed in this description: spherical harmonic functions and discrete panning functions. Practical methods for designing the set of reconstruction filters Li(f) and Ri(f) will be described in more detail. From the discussion which follows, it will be clear to a person of ordinary skill in the relevant art that other spatial functions can be used and that alternative techniques for producing the corresponding reconstruction filters are available.
The extraction of the interaural time delay difference, ITD(σp, φp), from the HRTF pair L(σp, φp, f) and R(σp, φp, f) is performed as follows.
Any transfer function H(f) can be uniquely decomposed into its all-pass component and its minimum-phase component as follows:
H(f)=exp(jφ(f))Hmin(f)
where φ(f), called the excess-phase function of H(f), is defined by
φ(f)=Arg(H(f))−Re(Hilbert(−Log|H(f)|)).
Applying this decomposition to the HRTFs L(σp, φp, f) and R(σp, φp, f), we obtain the corresponding excess-phase functions, φR(σp, φp, f) and φL(σp, φp, f), and the corresponding minimum-phase HRTFs, Lmin(σp, φp, f) and Rmin(σp, φp, f). The interaural time delay difference, ITD(σp, φp), can be defined, for each direction (σp, φp), by a linear approximation of the interaural excess-phase difference:
φR(σ,φ,f)−φL(σ,φ,f)≈2πfITD(σ,φ).
In practice, this approximation may be replaced by various alternative methods of estimating the ITD, including time-domain methods such as methods using the cross-correlation function of the left and right HRTFs or methods using a threshold detection technique to estimate an arrival time at each ear. Another possibility is to use a formula for modeling the variation of ITD vs. direction. For instance,
The delay-free HRTFs, L(σp, φp, f) and R(σp, φp, f), from which the reconstruction filters Li(f) and Ri(f) will be derived, can be identical, respectively, to the minimum-phase HRTF Lmin(σp, φp, f) and Rmin(σp, φp, f).
Whatever the method used to extract or model the interaural time delay difference from the measured HRTF, it can be regarded as an approximation of the interaural excess-phase difference φR(σ, φ, f)−φL(σ, φ, f) by a model function φ(σ, φ, f):
φR(σ,φ,f)−φL(σ,φ,f)≈φ(σ,φ,f).
It may be advantageous, in order to improve the fidelity of the 3-D audio reproduction according to the present invention, to correct for the error made in this phase difference approximation, by incorporating the residual excess-phase difference into the delay-free HRTFs L(σp, φp, f) and R(σp, φp, f) as follows:
L(f)=Lmin(f)exp(jφL(f)) and R(f)=Rmin(f)exp(jφR(f)),
where φL(f) and φR(f) satisfy
φR(f)−φL(f)=φR(f)−φL(f)−φ(σ,φ,f),
and either φL(f)=0 or φR(f)=0, as appropriate to ensure that the delay-free HRTFs L(σp, σp, f) and R(σp, σp, f) are causal transfer functions.
General Definition of Spherical Harmonics.
Of particular interest in the following description are the zero-order harmonic W and the first-order harmonics X, Y and Z defined earlier, as well as the second-order harmonics, U and V, and the third-order harmonics, S and T, defined below.
U(σ,φ)=cos2(φ)cos(2σ)
V(σ,φ)=cos2(φ)sin(2σ)
S(σ,φ)=cos3(φ)cos(3σ)
T(σ,φ)=cos3(φ)sin(3σ)
Advantages of spherical harmonics include:
As discussed above, a Soundfield microphone produces B format encoded signals. As such, a Soundfield microphone can be characterized by a set of spherical harmonic functions. Thus from
ITD(σ,φ)=tR(σ,φ)−tL(σ,φ)=d/c cos(φ)sin(σ),
where d is the distance between the microphones. If the ITD model provided in the encoder takes into account the diffraction of sound around the head or a sphere, the encoded signal and the recorded signal will differ in the value of the ITD for sounds away from the median plane. This difference can be reduced, in practice, by adjusting the distance between the two microphones to be slightly larger than the distance between the two ears of the listener.
The Binaural B Format recording technique is compatible with currently existing 8-channel digital recording technology. The recording can be decoded for reproduction over headphones through the bank of 8 filters Li(f) and Ri(f) shown on
The Binaural B Format offers the additional advantage that the set of four left or right channels can be used with conventional Ambisonic decoders for loudspeaker playback. Other advantages of using spherical harmonics as the spatial panning functions in carrying out the invention will be apparent in connection to multi-channel loudspeaker playback, offering an improved fidelity of 3-D audio reproduction compared to Ambisonic techniques.
For clarity, the derivation of the N reconstruction filters Li(f) will be illustrated in the case where the spatial panning functions gi(σp, φp) are spherical harmonics. However, the methods described are general and apply regardless of the choice of spatial functions.
The problem is to find, for a given frequency (or time) f, a set of complex scalars Li(f) so that the linear combination of the spatial functions gi(σp, φp) weighted by the Li(f) approximates the spatial variation of the HRTF L(σp, φp, f) at that frequency (or time). This problem can be conveniently represented by the matrix equation
L=GL,
where
The solution which minimizes the energy of the error is given by the pseudo inversion
L=(GTG)−1GTL,
where (GT G), known as the Gram matrix, is the N×N matrix formed by the dot products G(i, k)=GiT Gk of the spatial vectors. The Gram matrix is diagonal if the spatial vectors are mutually orthogonal.
Simplest case: the sampled spatial functions are mutually orthogonal => filters are derived by orthogonal projection of the HRTF on the individual spatial functions (dot product computed at each frequency). Example: 2-D reproduction with regular azimuth sampling. If sampled functions are not mutually orthogonal, multiply by inverse of Gram matrix to ensure correct reconstruction.
Even when the panning functions gi(σ, φ) are mutually ortogonal, as is the case with spherical harmonics, the vectors Gi obtained by sampling these functions may not be orthogonal. This happens typically if the spatial sampling is not uniform (as is often the case with 3-D HRTF measurements). This problem can be remedied by redefining the spatial dot product so as to approximate the continuous integral of the product of two spatial functions
<gi,gk>=1/(4π)∫σ∫σgi(σ,φ)gk(σ,φ)cos(φ)dσdφ
by
<gi,gk>=Σ{p=1, . . . P}gi(σp,φp)gk(σp,φp)dS(p)=GiTΔGk
where Δ is a diagonal P×P matrix with Δ(p, p)=dS(p) and dS(p) is proportional to a notional solid angle covered by the HRTF measured for the direction (σp, φp). This definition yields the generalized pseudo inversion equation
L=(GTΔG)−1GTΔL,
where the diagonal matrix Δ can be used as a spatial weighting function in order to achieve a more accurate 3-D audio reproduction in certain regions of space compared to others, and the modified Gram matrix (GT ΔG) ensures that the solution minimizes the mean squared error.
Additional possibility: project on a subset of the chosen set of spatial functions using above methods. Then project the residual error over other spatial functions (cf aes16). Example: to optimize fidelity of reconstruction in horizontal plane, project on W, X, Y first, and then project error on Z. Note that process can be iterated in more than 2 steps.
By combining the above techniques, it is possible, for a given set of spatial panning functions, to achieve control over chosen perceptual aspects of the 3-D audio reproduction, such as the front/back or up/down discrimination or the accuracy in particular regions of space.
An advantage of a recording mad in accordance with the invention over a conventional two-channel dummy head recording is that, unlike prior art encoded signals, binaural B format encoded signals do not contain spectral HRTF features. These features are only introduced at the decoding stage by the reconstruction filters Li(f). Contrary to a conventional binaural recording, a Binaural B Format recording allows listener-specific adaptation at the reproduction stage, in order to reduce the occurrence of artifacts such as front-back reversals and in-head or elevated localization of frontal sound events.
Listener-specific adaptation can be achieved even more effectively in the context of a real-time digital mixing system. Moreover, the technique of the present invention readily lends itself to a real-time mixing approach and can be conveniently implemented as it only involves the correction of the head radius r for the synthesis of ITD cues and the adaptation of the four reconstruction filters Li(f). If diffuse-field equalization is applied to the headphones and to the measured HRTF, and therefore to the reconstruction filters Li(f), the adaptation only needs to address direction-dependent features related to the morphology of the listener, rather than variations in HRTF measurement apparatus and conditions.
Definition: functions which minimize the number of non-zero panning weights for any direction: 2 weights in 2D and 3 weights in 3D. For each panning function, there is a direction where this panning function reaches unity and is the only non-zero panning function. Example given in
An advantage of discrete panning functions: fewer operations needed in encoding module (multiplying by panning weight and adding into the mix is only necessary for the encoding channels which have non-zero weights).
The projection techniques described above can be used to derive the reconstruction filters. Alternatively, it can be noted that each discrete panning function covers a particular region of space, and admits a “principal direction” (the direction for which the panning weight reaches 1). Therefore, a suitable reconstruction filter can be the HRTF corresponding to that principal direction. This will guarantee exact reconstruction of the HRTF for that particular direction. Alternatively, a combination of the principal direction and the nearest directions can be used to derive the reconstruction filter. When it is desired to design a 3D audio display system which offers maximum fidelity for certain directions of the sound, it is straightforward to design a set of panning functions which will admit these specific directions as principal directions.
When used in the topologies of
The binaural signal is decomposed as follows:
L(σ,φ,f)=LF(σ,φ,f)+LB(σ,φ,f)
where LF and LB are the “front” and “back” binaural signals, defined by:
LF(σ,φ,f)=0.5{[W(σ,φ)+X(σ,φ)][LW(f)+LX(f)]+Y(σ,φ) LY(f)+Z(σ,φ)LZ(f)}
LB(σ,φ,f)=0.5{[W(σ,φ)−X(σ,φ)][LW(f)−LX(f)]+Y(σ,φ)LY(f)+Z(σ,φ)LZ(f)}
It can be verified that LB=0 for (σ, φ)=(0, 0) and that LF=0 for (σ, φ)=(π, 0). The network of
The following notations are used in
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/US99/22259 | 9/24/1999 | WO | 00 | 1/9/2002 |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO00/19415 | 4/6/2000 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
4086433 | Gerzon | Apr 1978 | A |
5436975 | Lowe et al. | Jul 1995 | A |
5500900 | Chen et al. | Mar 1996 | A |
5521981 | Gehring | May 1996 | A |
5596644 | Abel et al. | Jan 1997 | A |
5757927 | Gerzon et al. | May 1998 | A |
5802180 | Abel et al. | Sep 1998 | A |
5809149 | Cashion et al. | Sep 1998 | A |
6259795 | McGrath | Jul 2001 | B1 |
6418226 | Mukojima | Jul 2002 | B2 |
6577736 | Clemow | Jun 2003 | B1 |
6628787 | McGrath et al. | Sep 2003 | B1 |
6990205 | Chen | Jan 2006 | B1 |