Canary Song Decoder: Transduction and Implicit Segmentation with ESNs and LTSMs

Abstract

Domestic canaries produce complex vocal patterns embedded in various levels of abstraction. Studying such temporal organization is of particular relevance to understand how animal brains represent and process vocal inputs such as language. However, this requires a large amount of annotated data. We propose a fast and easy-to-train transducer model based on RNN architectures to automate parts of the annotation process. This is similar to a speech recognition task. We demonstrate that RNN architectures can be efficiently applied on spectral features (MFCC) to annotate songs at time frame level and at phrase level. We achieved around 95% accuracy at frame level on particularly complex canary songs, and ESNs achieved around 5% of word error rate (WER) at phrase level. Moreover, we are able to build this model using only around 13 to 20 minutes of annotated songs. Training time takes only 35 seconds using 2 hours and 40 minutes of data for the ESN, allowing to quickly run experiments without the need of powerful hardware.

Create Efficient and Complex Reservoir Computing Architectures with ReservoirPy

“…Although existing since the beginning of the 2000s, RC techniques are less well-known compared to other RNN-based Deep Learning architectures like Long Short-Term Memory networks (LSTMs). In the meantime, they have been successfully applied to various tasks and problems (some are listed in this review by [25]) and even demonstrates state of the art performances for tasks such as chaotic timeseries forecasting [30] or sound processing [27]. It was shown that ESNs needed less data than LSTMs to obtain good performances while being trained in much less time (e.g.…”

“…It was shown that ESNs needed less data than LSTMs to obtain good performances while being trained in much less time (e.g. see [27]).…”