Académique Documents
Professionnel Documents
Culture Documents
be the EVD of the M × M covariance matrix. Then, U contains for various combinations of {ωk , Lk , Q} using the FFT-based im-
the M orthonormal eigenvectors of R, i.e., U = [ u1 · · · uM ] plementation described in [5]. Note that this optimization other-
and Λ is a diagonal matrix containing the corresponding eigenval- wise involves a multi-dimensional evaluation of the cost function
ues, λk , with λ1 ≥ λ2 ≥ . . . ≥ λM . Let GQ be formed from the over all combinations of {ωk }K K
k=1 and {Lk }k=1 . The number of
eigenvectors corresponding to the M − Q least significant eigen- sources can also be determined using this method for more than
values, i.e., one source, i.e., K > 1, by allowing Lk = 0 for the other sources.
ˆ ˜ For more on order estimation using the MUSIC orthogonality prin-
GQ = uQ+1 · · · uM . (9)
ciple and its performance, we refer the interested reader to [5]. For
The noise subspace GQ will then be orthogonal to Z, i.e., the special case where the number of harmonics Lk = L, ∀k, is
known and equal, i.e., Q = KL, the method reduces to the one-
ZH GQ = 0. (10) dimensional minimization over ω, i.e.,
This is the orthogonality principle of MUSIC and it can be used X
L
for finding model parameters and subspace ranks. In practice, this AH 2
k GQ F = aH (ωl)GQ GH
Q a(ωl), (17)
orthogonality will hold only approximately and can be measured l=1
using the Frobenius norm.
where the fundamental frequencies can be identified as the K deep-
est valleys in the cost function. The complexity of the proposed
3. MULTI-PITCH HARMONIC MUSIC estimator can be reduced by first finding the appropriate ranks of
the signal and noise subspaces using unconstrained frequencies.
Having introduced the covariance matrix model and the MUSIC This can be done using a simple special case of the proposed es-
orthogonality principle, we now proceed to present the proposed timator, where first it is assumed that Lk = 1, ∀k. Then, the
method. Estimates are obtained using MUSIC as the frequencies total number of sinusoids, and thereby the signal and noise sub-
minimizing the cost function space ranks, can be estimated as the number of sources K. The
harmonic MUSIC algorithm can then be applied given these sub-
ZH GQ 2F
J= , (11) space ranks, whereby the total number of different combinations
M Q(M − Q) of orders {Lk } is greatly reduced.
For a given or estimated Lk , the gradient of the cost function
with · F denoting the Frobenius norm and M Q(M − Q) being
(11) can be shown to be (for simplicity we here drop the scaling)
an order-dependent scaling. If this scaling is omitted, the estimator
can easily be observed to be biased. For more on MUSIC and its „ j ff«
∂J H ∂Ak
performance see, e.g., [12, 13]. The set of fundamental frequencies ∇J = 2 Re Tr AH k GQ GQ , (18)
∂ω0 ∂ω0
can be found as (see [5])
with Re(·) denoting the real value, the Schur-Hadamard prod-
ZH GQ 2F
{ω̂k }K
k=1 = arg min min (12) uct, and
{ωk }K {Lk }K M Q(M − Q) ∂Ak
k=1 k=1 = Yk Ak (19)
∂ω0
X
K
AH 2
k GQ F
= arg min min . (13) with 2 3
{ωk }K
k=1
{Lk }K
k=1
M Q(M − Q) 0 ··· 0
k=1
6 j ··· jLk 7
6 7
Yk = 6 .. .. .. 7. (20)
4 . . . 5
j(M − 1) · · · j(M − 1)Lk
Similarly, the Hessian can be derived to be 0.4
∂2J 0.35
∇2 J (21)
∂ω02
` ˘ H` ´ 0.3
= 2 Re Tr AHk GQ GQ Yk Yk Ak (22)
` ´H ` ´¯´
Frequency [Hz]
0.25
+ Yk Ak GQ GH Q Yk Ak . (23)
0.2
The gradient and the Hessian can be used for finding refined esti-
mates using standard methods. Here, we iteratively find a refined 0.15
estimate of the fundamental frequency using Newton’s method,
i.e., 0.1
(i+1) (i) ∇J
ω̂k = ω̂k − δ 2 , (24)
∇ J 0.05
with i being the iteration index and δ a small, positive constant,
0
which is found using approximate line search. The method is ini- 0 20 40 60 80 100
tialized for i = 0 using the coarse fundamental frequency estimate Segment
and order obtained from (15). As the order Lk and the signal sub-
space rank Q is kept fixed in the Newton method, only the poles Fig. 1. True (solid) and estimated (circles) fundamental frequen-
of the Vandermonde matrix Ak changes in each iteration. cies for two sources having unknown orders.
−2
10
HMUSIC
4. NUMERICAL RESULTS
CRLB
−4
10 vol. 11(7), pp. 609–612, July 2004.
[3] H. Li, P. Stoica, and J. Li, “Computationally efficient pa-
rameter estimation for harmonic sinusoidal signals,” Signal
10
−5 Processing, vol. 80, pp. 1937–1944, 2000.
[4] M. G. Christensen, S. H. Jensen, S. V. Andersen, and
A. Jakobsson, “Subspace-based fundamental frequency esti-
−6
mation,” in Proc. European Signal Processing Conf., 2004,
10 pp. 637–640.
100 200 300 400 500 600 700 800
N
[5] M. G. Christensen, A. Jakobsson, and S. H. Jensen, “Joint
high-resolution fundamental frequency and order estima-
Fig. 3. RMSE and CRLB as a function of the number of obser- tion,” IEEE Trans. on Audio, Speech and Language Pro-
vations, N , for P SN R = 40 dB and two sources having known cessing, Apr. 2006, submitted.
orders.
[6] R. Gribonval and E. Bacry, “Harmonic Decomposition of
Audio Signals with Matching Pursuit,” IEEE Trans. Signal
0
10
Processing, vol. 51(1), pp. 101–111, Jan. 2003.
HMUSIC [7] A. Klapuri and M. Davy, Eds., Signal Processing Methods
CRLB
−1 for Music Transcription, Springer, New York, 2006.
10
[8] A. Klapuri, “Multiple fundamental frequency estimation
−2 based on harmonicity and spectral smoothness,” IEEE Trans.
10
Speech and Audio Processing, vol. 11(6), pp. 804–816, 2003.
RMSE