Acoustic fingerprinting and graphical soundscapes

Acoustic fingerprinting is a technique that captures unique features of audio signals. For example, Shazam employs a spectrogram-based approach, converting audio into a visual representation and then identifying peaks on the spectrogram [1]. This fingerprint is matched against a vast database to identify the corresponding song. The method is robust in presence of noise, allowing accurate recognition of diverse audio sources in real-time. This approach is versatile, finding application in characterizing soundscapes. It has been successfully employed to evaluate FSC forest certification [2] and Neotropical oil palm landscapes [3].

Load required modules

from maad import sound, util, rois, features

Local maxima on spectrograms

Load the audio file, get the spectrogram, and compute the local maximas.

s, fs = sound.load('../../data/spinetail.wav')
Sxx, tn, fn, ext = sound.spectrogram(s, fs, nperseg=1024, noverlap=512)
Sxx_db = util.power2dB(Sxx, db_range=80)
peak_time, peak_freq = rois.spectrogram_local_max(
    Sxx_db, tn, fn, ext, min_distance=1, threshold_abs=-40, display=True)

Graphical soundscapes

If we compute the local maxima over multiple audio recordings from the same site, we can have a graphical representation of the most prominent spectro-temporal dynamics over a 24-hour window. To illustrate this, we will use 96 audio recordings that were collected in a temperate forest, and that are available here.

df = util.get_metadata_dir('../../data/indices')
df['time'] = df.date.dt.hour
gs = features.graphical_soundscape(
    data=df, time='time', threshold_abs=-80, target_fs=22000, verbose=False)
ax = features.plot_graph(gs)

This representation can be computed at various locations, and comparisons can be made among these sites for similarity. For further details on this approach, refer to sources [2] and [3]. It is important to note that the method implemented in scikit-maad is akin but not identical. In the original approach, peaks were calculated based on the mean spectrogram. In this version, we identify local maxima directly within the spectrogram.

References

[1] Wang, A. (2003, October). An industrial strength audio search algorithm. In Ismir (Vol. 2003, pp. 7-13).

[2] Campos‐Cerqueira, M., Mena, J. L., Tejeda‐Gómez, V., Aguilar‐Amuchastegui, N., Gutierrez, N., & Aide, T. M. (2020). How does FSC forest certification affect the acoustically active fauna in Madre de Dios, Peru?. Remote Sensing in Ecology and Conservation, 6(3), 274-285.

[3] Furumo, P. R., & Mitchell Aide, T. (2019). Using soundscapes to assess biodiversity in Neotropical oil palm landscapes. Landscape Ecology, 34, 911-923.