Academic
Publications
The INTERSPEECH 2010 paralinguistic challenge

The INTERSPEECH 2010 paralinguistic challenge,Björn Schuller,Stefan Steidl,Anton Batliner,Felix Burkhardt,Laurence Devillers,Christian A. Müller,Shrik

The INTERSPEECH 2010 paralinguistic challenge   (Citations: 7)
BibTex | RIS | RefWorks Download
Most paralinguistic analysis tasks are lacking agreed-upon evaluation procedures and comparability, in contrast to more 'traditional' disciplines in speech analysis. The INTERSPEECH 2010 Paralinguistic Challenge shall help overcome the usually low compatibility of results, by addressing three selected sub- challenges. In the Age Sub-Challenge, the age of speakers has to be determined in four groups. In the Gender Sub-Challenge, a three-class classification task has to be solved and finally, the Affect Sub-Challenge asks for speakers' interest in ordinal rep- resentation. This paper introduces the conditions, the Challenge corpora "aGender" and "TUM AVIC" and standard feature sets that may be used. Further, baseline results are given. Index Terms: Paralinguistic Challenge, Age, Gender, Affect
Cumulative Annual
View Publication
The following links allow you to view full publications. These links are maintained by other sources not affiliated with Microsoft Academic Search.
    • ...The signal-level audio feature set is based on the one used for the baseline results of the INTERSPPECH 2010 Paralinguistic Challenge [35]...
    • ...Thus, for future continuous dimensional affect prediction systems we should focus on the correlation coefficient as a main evaluation metric, as followed in the INTERSPEECH 2010 Paralinguistic Challenge [35]...

    Florian Eybenet al. String-based audiovisual fusion of behavioural events for the assessme...

    • ...We follow the official partitioning of the corpus as was used for the INTERSPEECH 2010 Paralinguistic Challenge [21]...

    Florian Eybenet al. Audiovisual classification of vocal outbursts in human conversation us...

    • ...In order to investigate spontaneous speech with non-restricted spoken content and natural expression of speaker state, it was decided to first include the TU Munich Audiovisual Interest Corpus (TUM AVIC) (Schuller et al. 2009 )a s recently featured in the INTERSPEECH 2010 Paralinguistic Challenge (Schuller et al. 2010a) in the experiments...

    Björn Schuller. Affective speaker state analysis in the presence of reverberation

    • ...Though, recent research efforts tend to the establishment [12,13] and utilization of a universally accepted setup [14,15]...

    Theodoros Kostoulaset al. Enhancing Emotion Recognition from Speech through Feature Selection

    • ...In this context, the organizers of the Interspeech 2010 Paralinguistic Challenge [23] defined an interest recognition task with unified system training and test conditions in order to make the recognition approaches developed by different researchers easily comparable...
    • ...In contrast to the baseline Paralinguistic Challenge recognition system that has been applied and evaluated in [23] and is based on acoustic features processed via unpruned REP-Trees, our proposed system also makes use of linguistic information obtained by automatic speech recognition (ASR) and exploits a self-learned amount of...
    • ...5 are based on the “TUM AVIC” corpus [20] which had also been used for the Affect Sub-Challenge of the Interspeech 2010 Paralinguistic Challenge [23]...
    • ...We will review the set of acoustic features that has been proposed in [23] in order to define a unified feature set that can be used for comparing the accuracy of different classification approaches (Sect...
    • ...4.1). Unlike the baseline interest predictor that had been introduced in [23] and is exclusively based on acoustic descriptors, the regression technique applied in this article also makes use of linguistic features and thus requires an ASR module recognizing spoken content and non-linguistic vocalizations such as laughing...
    • ...5 correspond to the baseline feature set of the Interspeech 2010 Paralinguistic Challenge [23]...
    • ...In conformance with [23], we chose the cross correlation (CC) between the ground truth level of interest and the predicted level of interest as evaluation criterion...
    • ...As an example, when evaluating a (‘dummy’) classifier that always predicts the mean of the training set ground truth labels, we obtain an MLE of 0.148 (which is only 0.002 below the MLE reported in [23]) while we get a CC of zero...
    • ...Table 4 Results for interest recognition as defined in the Affect Sub- -Challenge [23]: cross correlation obtained for different network architectures when using either acoustic (Ac.) or combined acoustic-linguistic (Ac...
    • ...+ Ling.) information; baseline results reported in [23 ]w hen applying unpruned REP-Trees with and without correlation-based feature selection (CFS)...
    • ...REP-Trees Yes 0.439 0.435 REP-Trees [23] No 0.421 0.423...
    • ...For comparison, also the Paralinguistic Challenge baseline result (CC of 0.421, obtained with unpruned REP-Trees in Random-Sub-Space meta-learning [23]) is shown in Table 4. The REP-Trees approach profits from feature selection via CFS but cannot compete with the BLSTM technique...

    Martin Wöllmeret al. Computational Assessment of Interest in Speech—Facing the Real-Life Ch...

Sort by: