MSc thesis project proposal

Intelligent sound recognition using spiking neural network

Project outside the university

Innatera
Innatera is a semiconductor company that develops microprocessors that are based on the architecture of the brain. These devices mimic the brain’s mechanisms for processing fast information streams from sensors, and enable complex, turn-key sensor analytics functionalities, with 10,000x higher performance per watt than conventional microprocessors. Innatera’s technology enables intelligence functionalities to be realized in devices at the extreme edge, and is a critical enabler for next-generation applications in the IoT, wearable, embedded, and automotive domains. Innatera is a spin-off of the Delft University of Technology, and is based on the university campus in Delft.

Project description The ability to recognize sound is an important requirement in modern electronic devices, especially since the introduction of intelligent assistants like the Amazon Alexa. However, this trend is also observed in other markets including industrial electronics, security, and automotive. Common use-cases built on sound recognition include:
  • Voice and spoken word recognition
  • Aggression/accident detection in crowded spaces
  • Speaker identification
  • Sound localization
  • Anomaly detection in machines
A large majority of such use-cases are passive in nature, i.e. involve continuous monitoring of the environment until the relevant event occurs. This means that the listening functions are continuously active (always-on), and the processing pipeline that interprets the resulting data stream is continuously running. When such use-cases are realized in power-limited devices (eg. wearables with a limited battery capacity, or industrial system within a small, thermally constrained housing), power dissipation of the sensing and processing functions becomes a critical factor. Notably, the viability of intelligent sensing concepts is predicated upon the power dissipation of the processing pipeline being maintained within a narrow envelope, typically under 10mW for wearables. Conventional signal processing approaches applied to these use-cases generally incur a power cost that lies outside of this envelope.

Objective In this thesis assignment, you will explore brain-inspired algorithms for carrying out sound recognition in cuttingedge, power-constrained application use-cases. The brain relies on a highly-efficient neural network – the spiking neural network – to process sensor information from the sense organs. In this assignment, you will explore and develop spiking neural network concepts for sound processing, and develop an algorithm that enables:
  1. Characterization of sounds in audio data (eg. from a microphone)
  2. Identification of characterized sounds
  3. Mitigation of noise and other undesirable characteristics
As part of the assignment, you will explore the design space for the problem, evaluate and benchmark state-of-theart solutions, develop an algorithm using a high-level simulation framework, and subsequently, potentially implement the developed algorithm on Innatera’s experimental hardware platform. This assignment is an exciting opportunity to participate in a highly innovative technology development, and offers a chance to shape how Innatera’s ground-breaking processors will address the needs of next-generation use-cases in the industry.

Location & Supervision This assignment will be carried out at Innatera Nanosystems in Delft, and will be performed under the supervision of dr. Richard Hendriks of the Circuits & Systems Group, TUDelft.

Contact

dr.ir. Richard Hendriks

Circuits and Systems Group

Department of Microelectronics

Last modified: 2020-02-17