Robust Real-Time Vowel Classification with an Echo State Network

In the field of reservoir computing echo state networks (ESNs) and liquid state machines (LSMs) are the most commonly used networks. Comparative studies on these reservoirs identify the LSM as the network that yields the highest performance for speech recognition. But LSMs are not always usable in a real-time setting due to the computational costs of a large reservoir with spiking neurons. In this paper a vowel classification system is presented which consists of an ESN which processes cochlear filtered audio. The performance of the system is tested on a vowel classification task using different signal-to-noise ratios (SNRs). The usefulness of this method is measured by comparing it to formant based vowel classification systems. Results show that this ESN based system can get a performance similar to formant based vowel classification systems on the clean dataset with only a small reservoir and even outperforms these methods on the noisy dataset.