Re-annotation of cough events in the AMI corpus

TL;DR

This paper re-annotates cough events in the AMI corpus to create a reliable, publicly available database for developing and evaluating machine learning algorithms for audio cough detection, addressing privacy concerns.

Contribution

It introduces a new re-annotation methodology and provides a publicly accessible, annotated cough sound database based on the AMI corpus.

Findings

Re-annotated 1369 cough events in the AMI corpus.

Developed a MATLAB GUI for efficient annotation.

Made the cough annotations and tool publicly available.

Abstract

Cough sounds act as an important indicator of an individual's physical health, often used by medical professionals in diagnosing a patient's ailments. In recent years progress has been made in the area of automatically detecting cough events and, in certain cases, automatically identifying the ailment associated with a particular cough sound. Ethical and sensitivity issues associated with audio recordings of coughs makes it more difficult for this data to be made publicly available. However, without the public availability of a reliable database of cough sounds, developments in the area of audio event detection are likely to be hampered. The purpose of this paper is to spread awareness of a database containing a large amount of naturally occurring cough sounds that can be used for the implementation, evaluation, and comparison of new machine learning algorithms that allow for audio event detection associated with cough sounds. Using a purpose built GUI designed in MATLAB, the re-annotation procedure followed a reusable methodology that allowed for quick and efficient importing and marking of audio signals, resulting in a re-annotated version of the Augmented Multi-party Interaction (AMI) corpus' cough location annotations, with 1369 individual cough events. All cough annotations and the re-annotation tool are made available for download and public use.

Tables

Table 1. Table 1: Number of cough event annotations

Number of cough annotations
Original annotations	Re-annotated list
1116	1369

Full text

Re-annotation of cough events

in the AMI corpus

Paul Leamy, Ted Burke, Damon Berry, David Dorran

Biomedical Research Group

Technological University Dublin

(17-18 June 2019)

1 Introduction

Cough sounds act as an important indicator of an individual’s physical health [1], often used by medical professionals in diagnosing a patient’s ailments [2]. In recent years progress has been made in the area of automatically detecting cough events and, in certain cases, automatically identifying the ailment associated with a particular cough sound. Ethical and sensitivity issues associated with audio recordings of coughs make it more difficult for this data to be made publicly available. Without the public availability of a reliable database of cough sounds, developments in the area of audio event detection are likely to be hampered [3]. The purpose of this paper is to spread awareness of a database [4] containing a large amount of naturally occurring cough sounds that can be used for the implementation, evaluation, and comparison of new machine learning algorithms that allow for audio event detection associated with cough sounds. Using a purpose built GUI designed in MATLAB, a re-annotated version of the Augmented Multi-party Interaction (AMI) corpus’ cough location annotations was produced with 1369 individual cough events. All cough annotations and the re-annotation tool are available for download and public use [5].

2 The AMI corpus

The AMI corpus data is publicly available under the Creative Commons Attribution 4.0 license agreement, and contains audio recordings of indoor meetings taken in an office environment.

The audio recordings are accompanied by annotations of a large number of audio events throughout the entire corpus. Audio recordings were captured from multiple microphone locations, with an example of one meeting room shown in a Figure 1.

2.1 Re-annotation procedure

Following an initial analysis of the annotations, it was found that a number of issues existed relating to the original cough event annotations in the AMI corpus including. These include incorrect start and end times, and successive coughs annotated as single coughs, examples of which are illustrated in Figure 2.

The re-annotation criteria for the cough events is illustrated in Figure 3.

The task of importing the audio, and the recording of the new time-stamps was carried out using a purpose built GUI, as shown in Figure 4.

3 Results

Following the re-annotation procedure, a new set of time-stamps corresponding to the start and end times of single cough events was produced. The total number of single cough events before and after the re-annotation procedure is displayed in Table 1.

4 Conclusions

Justification for the re-annotation of the cough locations in the AMI corpus came following the discovery of a number of issues relating to the original annotations. In the area of cough sound analysis and detection the presence of a reliably annotated database will be a useful addition particularly in the area of machine learning. This database will be useful for progressing the development and testing of new audio event detection algorithms concerned with cough sounds. The re-annotated version of the cough location annotations have been made publicly available along with the GUI used in the re-annotation procedure [5]. A consistent approach to labelling of start and end times of the individual cough events was used, and as a result the potential for the AMI corpus to become a more reliable and reusable source of annotated cough recordings has increased.

4.1 References