Role of non-linear data processing on speech recognition task in the framework of reservoir computing

Area of Science:

• Neuromorphic computing
• Artificial intelligence
• Signal processing

Background:

• Reservoir computing neural networks are key for testing neuromorphic hardware.
• Automatic speech recognition (ASR) is a standard benchmark task.
• Acoustic transformations in ASR can obscure hardware performance.

Purpose of the Study:

• To quantify and separate the contributions of acoustic transformations and neuromorphic hardware to ASR success rate.
• To evaluate the role of non-linearity in acoustic feature extraction.
• To establish a benchmark for comparing different neuromorphic hardware.

Main Methods:

• Quantifying and separating contributions of acoustic transformations and hardware.
• Analyzing non-linearity in acoustic feature extraction.
• Benchmarking reservoir computing devices against acoustic transformations alone.
• Experimentally and numerically evaluating magnetic nano-oscillator hardware.

Main Results:

• Non-linear acoustic transformations significantly impact feature extraction.
• Reservoir computing devices offer a measurable gain in word success rate over acoustic transformations alone.
• The proposed method provides an appropriate benchmark for comparing neuromorphic hardware.

Conclusions:

• Acoustic transformation non-linearity is crucial for effective feature extraction in ASR.
• The developed benchmark accurately assesses neuromorphic hardware performance.
• Magnetic nano-oscillator hardware performance is quantified under different acoustic transformations.