Keyword Spotting System using Low-complexity Feature Extraction and Quantized LSTM

Archive ouverte : Communication dans un congrès

Hérissé, Kévin | Larras, Benoit | Frappé, Antoine | Kaiser, Andreas

Edité par HAL CCSD

International audience. Long Short-Term Memory (LSTM) neural networks offer state-of-the-art results to compute sequential data and address applications like keyword spotting. Mel Frequency Cepstral Coefficients (MFCC) are the most common features used to train this neural network model. However, the complexity of MFCC coupled with highly optimized machine learning neural networks usually makes the MFCC feature extraction the most power-consuming block of the system. This paper presents a low complexity feature extraction method using a filter bank composed of 16 channels with a quality factor of 1.3 to compute a spectrogram. It shows that we can achieve an 89.45% accuracy on 12 classes of the Google Speech Command Dataset using an LSTM network of 64 hidden units with weights and activation quantized to 9 bits and inputs quantized to 8 bits.

• Contient
• Sujet
• Description
• Oeuvre