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.

Consulter en ligne

Suggestions

Du même auteur

Mixed-Signal In-Memory Multi-bit Matrix-Vector Multiplication | Hérissé, Kévin

Mixed-Signal In-Memory Multi-bit Matrix-Vector Multiplication

Archive ouverte: Communication dans un congrès

Hérissé, Kévin | 2021-06-08

poster. International audience. The applications for artificial intelligence are wide and cover multiple domains including industry, health, home automation, consumer electronics, automotive, and smart cities. Appli...

Mixed-Signal In-Memory Multi-bit Matrix-Vector Multiplication | Hérissé, Kévin

Mixed-Signal In-Memory Multi-bit Matrix-Vector Multiplication

Archive ouverte: Poster de conférence

Hérissé, Kévin | 2021-10-13

International audience

Instantaneous power estimation in a continuous time digital signal processing system | González Santos, Ángel de Dios

Instantaneous power estimation in a continuous time digital signal processi...

Archive ouverte: Brevet

González Santos, Ángel de Dios | 2021-08-19

N° de priorité : US202016789589 20200213. The present invention generally relates to the esti mation of instantaneous power of a signal and, in particular, to a circuit for use in a continuous time digital signal pr...

Du même sujet

Essential math for data science : take control of your data with fundamental linear algebra, probability, and statistics / Thomas Nield | Nield, Thomas. Auteur

Essential math for data science : take control of your data with fundamenta...

Livre | Nield, Thomas. Auteur | 2022

To succeed in data science you need some math proficiency. But not just any math. This common-sense guide provides a clear, plain English survey of the math you'll need in data science, including probability, statistics, hypothesi...

An Investigation on Inherent Robustness of Posit Data Representation | Alouani, Ihsen

An Investigation on Inherent Robustness of Posit Data Representation

Archive ouverte: Communication dans un congrès

Alouani, Ihsen | 2021-02-20

International audience. As the dimensions and operating voltages of computer electronics shrink to cope with consumers' demand for higher performance and lower power consumption, circuit sensitivity to soft errors i...

Lower Voltage for Higher Security: Using Voltage Overscaling to Secure Deep Neural Networks | Islam, Shohidul

Lower Voltage for Higher Security: Using Voltage Overscaling to Secure Deep...

Archive ouverte: Communication dans un congrès

Islam, Shohidul | 2021-11-01

International audience. Deep neural networks (DNNs) are shown to be vulnerable to adversarial attacks-- carefully crafted additive noise that undermines DNNs integrity. Previously proposed defenses against these att...

Improved Salp Swarm Optimization Algorithm: Application in Feature Weighting for Blind Modulation Identification | Ben Chaabane, Sarra

Improved Salp Swarm Optimization Algorithm: Application in Feature Weightin...

Archive ouverte: Article de revue

Ben Chaabane, Sarra | 2021-08

International audience. In modulation identification issues, like in any other classification problem, the performance of the classification task is significantly impacted by the feature characteristics. Feature wei...

[Review] Intelligent on-demand design of phononic metamaterials | Jin, Yabin

[Review] Intelligent on-demand design of phononic metamaterials

Archive ouverte: Article de revue

Jin, Yabin | 2022-01-25

International audience. With the growing interest in the field of artificial materials, more advanced and sophisticated functionalities are required from phononic crystals and acoustic metamaterials. This implies a ...

Breaking (and Fixing) Channel-based Cryptographic Key Generation: A Machine Learning Approach | Alouani, Ihsen

Breaking (and Fixing) Channel-based Cryptographic Key Generation: A Machine...

Archive ouverte: Communication dans un congrès

Alouani, Ihsen | 2022-08-31

International audience. Several systems and application domains are under-going disruptive transformations due to the recent breakthroughs in computing paradigms such us Machine Learning and commu-nication technolog...

Chargement des enrichissements...