Timbre is a key perceptual feature that allows discrimination between different sounds. Timbral sensations are highly dependent on the temporal evolution of the power spectrum of an audio signal. In order to quantitatively characterize such sensations, the shape of the power spectrum has to be encoded in a way that preserves certain physical and perceptual properties. Therefore, it is common practice to encode short-time power spectra using psychoacoustical frequency scales. In this paper, we study and characterize the statistical properties of such encodings, here called timbral code-words. In particular, we report on rank-frequency distributions of timbral code-words extracted from 740 hours of audio coming from disparate sources such as speech, music, and environmental sounds. Analogously to text corpora, we find a heavy-tailed Zipfian distribution with exponent close to one. Importantly, this distribution is found independently of different encoding decisions and regardless of the audio source. Further analysis on the intrinsic characteristics of most and least frequent code-words reveals that the most frequent code-words tend to have a more homogeneous structure. We also find that speech and music databases have specific, distinctive code-words while, in the case of the environmental sounds, this database-specific code-words are not present. Finally, we find that a Yule-Simon process with memory provides a reasonable quantitative approximation for our data, suggesting the existence of a common simple generative mechanism for all considered sound sources.

Our previous research showed promising results when transferring features learned from speech to train emotion recognition models for music. In this context, we implemented a denoising autoencoder as a pretraining approach to extract features from speech in two languages (English and Mandarin). From that, we performed transfer and multi-task learning to predict classes from the arousal-valence space of music emotion. We tested and analyzed intra-linguistic and cross-linguistic settings, depending on the language of speech and lyrics of the music. This paper presents additional investigation on our approach, which reveals that: (1) performing pretraining with speech in a mixture of languages yields similar results than for specific languages - the pretraining phase appears not to exploit particular language features, (2) the music in Mandarin dataset consistently results in poor classification performance - we found low agreement in annotations, and (3) novel methodologies for representation learning (Contrastive Predictive Coding) may exploit features from both languages (i.e., pretraining on a mixture of languages) and improve classification of music emotions in both languages. From this study we conclude that more research is still needed to understand what is actually being transferred in these type of contexts.

In this paper we present a method to search for environmental sounds in large unstructured databases of user-submitted audio, using a general sound events taxonomy from ecological acoustics. We discuss the use of Support Vector Machines to classify sound recordings according to the taxonomy and describe two use cases for the obtained classification models: a content-based web search interface for a large audio database and a method for segmenting field recordings to assist sound design.