Towards efficient keyword spotting using spike-based time difference encoders