After running the current version of MSAF with the following recommended parameters, the HitRate_t3F of the Structural Features algorithm on the BeatlesTUT dataset only obtains approximately 0.57. The published msaf-experiments data shows approximately 0.63. Other metrics for BeatlesTUT and SALAMI datasets are similarly lower.
Recommended parameters:
Feature parameters:
Type: PCP
sample_rate: 11025
n_fft: 2048
hop_size: 512
pcp.bins: 87
pcp.norm: np.inf
pcp.f_min: 27.5
pcp.n_octaves: 7
Algorithm configuration:
M_gaussian: 23
m_embedded: 3
Mp_adaptive: 28
offset_thres: 0.05
k_nearest: 0.04
bound_norm_feats: np.inf
By increasing the sampling rate to 22050 I get a HitRate_t3F of 0.59.
I am currently evaluating the MSAF v0.1.0 release with the minimum allowable package versions specified in the setup.py file.
After running the current version of MSAF with the following recommended parameters, the HitRate_t3F of the Structural Features algorithm on the BeatlesTUT dataset only obtains approximately 0.57. The published msaf-experiments data shows approximately 0.63. Other metrics for BeatlesTUT and SALAMI datasets are similarly lower.
Recommended parameters:
Feature parameters:
Type: PCP
sample_rate: 11025
n_fft: 2048
hop_size: 512
pcp.bins: 87
pcp.norm: np.inf
pcp.f_min: 27.5
pcp.n_octaves: 7
Algorithm configuration:
M_gaussian: 23
m_embedded: 3
Mp_adaptive: 28
offset_thres: 0.05
k_nearest: 0.04
bound_norm_feats: np.inf
By increasing the sampling rate to 22050 I get a HitRate_t3F of 0.59.
I am currently evaluating the MSAF v0.1.0 release with the minimum allowable package versions specified in the setup.py file.