Figure 1. Overall schema of sound classification module

Le systeme developpe dans le cadre du projet CUIDADO, doit permettre la classification automatique des sons a partir de definition de classes prealable, ainsi que l'apprentissage de ces classes en lignes (par definition de l'utilisateur). Pour cela, un systeme complet d'apprentissage/evaluation a ete creer:

• Extraction des descripteurs
• Apprentissage des modeles de classes
1. selection automatique des meilleurs descripteurs,
2. transformation des descripteurs,
3. creation des modeles de classes
• Evaluation: mesure de l'appartenance d'un son donne a une classe definie

Figure 2. Features extraction

For each sound, a large set of sound descriptors are extracted including

• temporal shape descriptors
  (attack-time, increaqse/decrease time, modulation, ...)
• energy descriptors,
• spectral descriptors
  (centroid, spread, skewness, kurtosis, roll-roff, ...),
• harmonic descriptors
  (fundamental frequency, noisiness, inharmonicity, odd-to-even energy ratio, deviation, ...),
• perceptual features
  (loudness, specific loudness, roughness, ...).

The evolution along time of a specific features is then modeled by a temporal modeling modules.

Each sound is then represented by a feature vector.

Figure 3. Feature selection

In order to determine the most appropriate features to describe a specific taxonomy, a "feature selection" module is used.

1. This module select the best features according to the value of the Between class inertia to the Total inertia. The largest this ratio is, the most discriminant is the feature.
2. The whole feature space is then projected on the first selected feature (the one with the largest ratio value) and the process repeated for the selection of the next value.

Figure 4. Instrument sounds taxonomy

Six different databases have been used for the evaluation, for a total of 4163 sounds, covering 27 instrument classes. Three different classes taxonomy have been used:

• T1: sustained/non-sustained sounds,
• T2: instrument families,
• T3: instrument names.

Figure 5. Confusion matrix

The best results are obtained using the

• hierarchical gaussian/KNN classifiers.
• hierarchical KNN classifiers.

Recognition rates for 27 instruments:

• at T1: 99%,
• at T2: 84%,
• at T3: 64%.