IEEE Signal Processing Cup 2024

IEEE Signal Processing Cup at IEEE ICASSP 2024
ROBOVOX: Far-Field Speaker Recognition by a Mobile Robot

IEEE ICASSP 2024 Website | Monday, 15 April 2024 | 2024 SP Cup Official Document

[Sponsored by the MathWorks and IEEE Signal Processing Society]

Introduction

A speaker recognition system authenticates the identity of claimed users from a speech utterance. For a given speech segment called enrollment and a speech segment from a claimed user, the speaker recognition system will determine automatically whether both segments belong to the same speaker or not. The state-of-the-art speaker recognition systems mainly use Deep Neural Networks (DNN) to extract fixed-length speaker discriminant representations called speaker embeddings. The decision to accept or reject a speaker will be made by comparing speaker embeddings. The DNN-based speaker verification systems perform well in general, but there are some challenges that reduce their performance dramatically. Far-field speaker recognition is among the well-known challenges facing speaker recognition systems. The far-field challenge is intertwined with other variabilities such as noise and reverberation. Two main categories of speaker recognition systems are text-dependent speaker recognition and text-independent speaker recognition. In a text-dependent speaker recognition system, the speaker’s voice is recorded from predefined phrases, while, in text-independent speaker recognition, there is no constraint on the content of the spoken dialogue. The task of the IEEE Signal Processing Cup 2024 is text-independent far-filed speaker recognition under noise and reverberation for a mobile robot.

Task Description

The Robovox challenge is concerned with doing far-field speaker verification from speech signals recorded by a mobile robot at variable distances in the presence of noise and reverberation. Although there are some benchmarks in this domain such as VoiCes and FFSVC, they don’t cover variabilities in the domain of robotics such as the robot’s internal noise and the angle between the speaker and the robot. The VoiCes dataset is replayed speech recorded under different acoustical noises. A main drawback of the VoiCes is that it was recorded from played signals whereas our dataset is recorded with people speaking in noisy environments. The FFSVC is another far-field speaker recognition benchmark. However, these benchmarks helped the community significantly, we are introducing a new benchmark for far-field speaker recognition systems in order to address some new aspects. Firstly, our goal is to perform speaker recognition in a real application for the domain of mobile robots. In this domain, there are other variabilities that have not been addressed in previous benchmarks: the robot’s internal noise and the angle between the speaker and the robot. Furthermore, the speech signal has been recorded for different distances between the speaker and the robot. In the proposed challenge the following variabilities are present:

• Ambient noise leading to low signal-to-noise ratios (SNR): The speech signal is distorted with noise from fans, air conditioners, heaters, computers, etc.
• Internal robot noises (robot activators): The robot’s activator noise reverberates on the audio sensors and degrades the SNR.
• Reverberation: The phenomena of reverberation due to the configuration of the places where the robot is located. The robot is used in different rooms with different surface textures and different room shapes and sizes.
• Distance: The distance between the robot and speakers is not fixed and it is possible for the robot to move during the recognition.
• Babble noise: The potential presence of several speakers speaking simultaneously.
• Angle: The angle between speakers and the robot’s microphones In this challenge, two tracks will be proposed:
• Far-field single-channel tracks: In this task, one channel is used to perform the speaker verification. The main objective is to propose novel robust speaker recognition pipelines to tackle the problem of far-field speaker recognition in the presence of reverberation and noise.
• Far-field multi-channel tracks: In this task, several channels are used to perform speaker verification. The main objective is to develop algorithms that improve the performance of multi-channel speaker verification systems under severe noise and reverberation.

Full technical details, dataset(s), evaluation metrics, and all other pertinent information about the competition is located in the 2024 SP Cup Official Document (above).

Important Dates

• Challenge Announcement: November 2023
• Team Registration Deadline: 12 January 2024 - Registration Link
• Final Submission of Team's Work Deadline: 5 February 2024 [Submit Team's Work]
• Announcement of 3 Finalists Teams: 14 February 2024
• Final Competition at IEEE ICASSP 2024: 14-19 April 2024

Registration and Important Resources

Official SP Cup Team Registration

• All teams MUST be registered through the official competition registration system before the deadline  in order to be considered as a participating team. Teams also MUST acknowledge, agree to the SPS Student Terms and Conditions, and meet all eligibility requirements at the time of team registration as well as throughout the competition.
• Register your team for the 2024 SP Cup before 12 January 2024 and submit work before 5 February 2024 at the following link: 

Finalist Teams

GRAND PRIZE - Team: "IITH"
Indian Institute of Technology Hyderabad
Supervisor: Sri Rama Murty Kodukula
Tutor: Sreekanth Sankala
Students: Tejadhith Sankar, Atharv Ramesh, Nair Vaideeswaran A P, Himanshu Kumar Gupta, Anirudh Srinivasan

FIRST RUNNER-UP-Team: "HYU ASML"
Hanyang University, Seoul, Republic of Korea
Supervisor: Joon-Hyuk Chang
Tutor: Jeong-Hwan Choi
Students: Hee-Jae Lee, Hyun-Soo Kim, Seyun Ahn, Gaeun Kim

SECOND RUNNER-UP - Team: "Pseudo Spectrum"
Rajshahi University of Engineering and Technology
Supervisor: Mohammod Abdul Motin
Tutor: Md Abdur Raiyan
Students: A Md Anik Hasan, Sapnil Sarker Bipro, Muhammad Sudipto Siam Dip
 

Request Complimentary MATLAB

• MathWorks, Inc. continued to support the IEEE SP Cup. Participating students are encouraged to download the complimentary MathWorks Student Competitions Software for use in the competition at the MathWorks’ SP Cup webpage.
• Please click "Request Software" on the website, fill in the application form, and then submit the form. You will receive an email within 3-5 working days after submission. The email will inform you of the software download, installation, and activation steps if the request is approved.
• More technical resources such as videos, examples and documentations can be found at the MathWorks’ SP Cup webpage.

Contacts

Competition Organizers (technical, competition-specific inquiries): Mohammad MOHAMMADAMINI

SPS Staff (Terms & Conditions, Travel Grants, Prizes): Jaqueline Rash, SPS Membership Program and Events Administrator

SPS Student Services Committee: Angshul Majumdar, Chair

Sponsors

This competition is sponsored by the IEEE Signal Processing Society and MathWorks: