Advancing Temporal Spike Encoding for Efficient Speech Recognition

Traditional artificial neural models have demonstrated considerable promise, but they frequently lack the intricate temporal dynamics inherent to biological neural systems. Standard models utilize continuous values for information transmission, whereas biological neurons communicate predominantly through discrete spike events. This study investigates the encoding of auditory data into spike sequences, with a particular focus on two techniques: Send on Delta (SOD) and SinSpike. Using these methods, speech signal was transformed into spiking representations, which were then evaluated in a series of experiments. As the link between external auditory stimuli and neural responses, the development of effective spike encoding techniques is crucial for advancing both our theoretical understanding and the practical applications of spiking neural networks in audio processing tasks. We show that efficient encoding methods yield improved speech recognition task compared to the use of a neuromorphic sensor.