Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This work is based on unsupervised classification in DMM models, the application of which was developed by one of the authors in reference [9]. Data on the daily life of a person living alone, are obtained from the SINS database, used for detecting daily activities in a domestic environment using a network of acoustic sensors. Some points regarding the integration of the two methodologies need clarification.

-First, a diagram outlining each step of the proposed method is necessary to visualize the integration of the data (and its modifications) with the proposed classifier.

-The authors state that the "absence" class is not included in their classifier because it would be easy to integrate it with PIR sensors. This would mean integrating another system in addition to the microphones. Were experiments conducted considering this class with only the acoustic signals?

-Further explanation is needed regarding the modifications made to the original data, specifically the discretization and beamforming:

1. a) How does the residual vector quantizer work and the effects in the classifier?
2. b) How the entire dataset is expanded, such that each original audio segment corresponds to three monaural audio recordings? How are you sure that the used software is really generating signals like real audio signals (bouncing in multiple directions)?
3. c) The procedure: This residual is subsequently projected onto the closest entry in a second codebook, and a new residual is again computed. By repeating this process eight times, a sequence of eight codes representing the input embedding is obtained. Why is this done like that? It is not clear.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents a method for recognizing daily activities from household sounds. The main drawbacks of the paper consist in the evaluation part:

• The presence of other environmental sounds (TV, noise made by dropped objects, etc) are acknowledged but not explicitly modeled or mitigated in the experiments.
• The authors claim that no explicit procedure for selection is required, but performance still depends on term frequency ratio thresholds and large preset values, which are manually chosen and not learned.
• The approach depends on microphone array beamforming, which may be impractical or cost-prohibitive in real home deployments.
• The experiments rely on the DCASE2018 dataset, which contains recordings from only one individual (young male). This severely limits the generalizability of the method, especially given the paper’s target application to elderly living alone.
• Comparisons are mostly restricted to K-Means and Spectral Clustering. More recent unsupervised or self-supervised audio representation learning methods must be considered.
• Gibbs sampling for multi-stream DMMs may be computationally expensive, but runtime, convergence behavior, and scalability are not discussed.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

All recommendations and observations were addressed.

Author Response

Comments 1: All recommendations and observations were addressed.

Response 1: We sincerely appreciate your understanding of our initial work on unsupervised learning of activity recognition models from daily-life sounds. Regarding your previous comments on the exclusion of the activity “absence”, we have added the experimental results obtained by performing clustering without removing acoustic words appearing in the activity “absence” to Figure 9. In Section 5.2, we also clarify the performance improvement achieved when these words are removed. Additional clarifications We have added a paragraph summarizing the key contributions of this study in the Introduction section. We have revised the first paragraph of Section 6, which acknowledges that the results presented in this paper are based on data from a single participant—who is also a younger individual—and therefore cannot be readily generalized. We have also reviewed the sixth paragraph of Section 6, which explains how the proposed method suppresses noise and discusses the remaining challenges. Furthermore, we have revised the seventh paragraph of Section 6, which describes several system design issues relevant to developing a practical implementation, including how to determine the threshold for the term frequency ratio, how to design the input stream, and how many codebooks should be employed to generate acoustic words.

 

Reviewer 2 Report

Comments and Suggestions for Authors

Some of my comments were addressed, but some of them need more explanations:

-possibility of existing soundsin the environment can influence the recognised activity. In order to prove the noise reduction, a set of experiments must be provided

-the evaluation made only on a dataset with activities performed by only one young person can't be used for generalizing the results. In this case - the variations in household sounds caused by age-related declines in physical function are relatively minor in the context of activity recognition, and must be proven by experimental results.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf