Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents the design, implementation and preliminary evaluation of an assistive system that combines a mobile application for older adults with a companion web platform for caregivers, aimed at supporting structured speech practice outside clinical settings. The authors frame their main engineering contribution around a deterministic, rule-based workflow that orchestrates audio capture, cloud-based speech recognition, pronunciation matching, immediate multimodal feedback, and controlled progression across difficulty levels. 

The proposal is interesting and addresses a real and underexplored niche and the integration of caregivers as first-class actors in the architecture is a genuine strength. However, I believe there are still a number of points the authors should cover before the manuscript is ready. Below are my recommendations, framed as questions that I think deserve a careful response in the revision:

1. The manuscript repeatedly describes the proposal as an "algorithm" and as the "core computational contribution", yet what is presented in Figure 2 and Table 1 is essentially a control-flow loop wrapping an external ASR call, a string-matching comparison, and an if/else decision. Could the authors clarify in what sense this constitutes a novel algorithmic contribution as opposed to an application workflow or orchestration pattern? It would strengthen the paper considerably to either (a) formalize the matching/decision rule more rigorously, or (b) reposition the contribution honestly as a system architecture paper with an embedded decision procedure, and adjust the framing accordingly.

2. The pronunciation evaluation is described as based on "predefined matching rules" and "deterministic decision rules", but nowhere in the manuscript are these rules specified. Is the comparison a strict string equality between ASR output and target word? Is any normalization applied? Is there a similarity threshold? Without this information the algorithm is not reproducible, which directly contradicts one of the claims in the abstract. I strongly recommend including pseudocode at the rule level, not just at the workflow level.

3.  The state-of-the-art section would benefit from a broader and more recent grounding in user-centered design and behavior-change technologies aimed at older adult populations, since this is precisely where the authors situate their contribution. As it stands, the discussion of accessibility heuristics and motivational mechanisms for this demographic leans heavily on a small set of voice-assistant studies, and the gamification component is supported by only a couple of references that don't really engage with the engineering literature on technology-mediated behavior change. I would suggest the authors complement the related-work section with sources such as Petrovcic (2018), who provide a structured taxonomy of mobile-phone usability problems for older adults [Int. J. Hum.-Comput. Interact., 34(3), 251–264]; Pak & McLaughlin (2011), Designing Displays for Older Adults, which remains a foundational reference for sensory and cognitive accommodations in interfaces; and the recent review by Del-Valle-Soto et al. (2024), "A Comprehensive Review of Behavior Change Techniques in Wearables and IoT: Implications for Health and Well-Being" (Sensors, 24(8), 2429), which offers an engineering-oriented synthesis of feedback, gamification, personalization and self-monitoring as behavior-change techniques mediated by technology, directly relevant to the motivational logic the authors embed in their system and useful for justifying which gamified elements are reasonable to include for this user group and why. These would help anchor the design decisions in a more solid theoretical framework rather than in a series of ad-hoc citations.

4. The reported recognition accuracy of 90% over only 50 attempts is, frankly, a weak empirical basis on which to make claims about deployment readiness, and the authors themselves acknowledge this, but then proceed to make precisely such claims in Section 4.3 and the Discussion. How many distinct speakers does this number aggregate? How many distinct target words? Was it computed per-attempt or per-utterance? 

5. Figure 9 raises serious questions about the experimental protocol. The curve begins at 100% for roughly the first 15 attempts, then collapses abruptly to 82.5%, before slowly climbing back to 90%. A genuine cumulative success rate over independent attempts cannot stay flat at exactly 100% for 15 trials and then drop in a single step, that pattern is consistent either with a synthetic/idealized data series or with a highly non-random ordering of attempts. Could the authors clarify how attempts were ordered, whether the data is real or simulated, and whether per-speaker breakdowns are available? As it stands, this figure undermines rather than supports the reliability claim.

6. The cloud-based ASR is treated as a black box. Which provider/model was actually used? This is not a minor detail: latency, accuracy, language support for Ecuadorian Spanish, and even the legal/ethical handling of older adults' voice recordings depend critically on it. The Discussion mentions network-dependent delays and asynchronous processing, but no measurements are provided that decompose the 1.82 s latency into capture / upload / ASR / response phases. A simple latency breakdown table would substantially strengthen the technical evaluation.

7. The usability evaluation involves five dyads and five caregivers , which is at the lower bound of what is defensible and is justified. However, the SUS score reported shows remarkably low variability for an instrument administered to five independent participants on a 0–100 scale. Could the authors confirm whether participants completed the questionnaire independently and without discussion, and discuss the risk of acquiescence bias and social desirability, particularly when caregivers evaluate a tool built for them and the older adults under their care, by researchers physically present in the room? Some reflection on this threat to validity is warranted.

8. The Discussion repeatedly emphasizes that the system is "latency-aware" and that its architecture explicitly addresses interaction timing, but this is asserted rather than demonstrated. Beyond reporting an average of 1.82 s, what design mechanisms in the implementation actively manage latency? Is there local fallback when the network degrades? Is there a barge-in timeout calibrated to older-adult pause patterns? If none of these exist, I would recommend softening the framing on latency awareness in the Discussion and Conclusions.

9. A more minor but pervasive issue: the manuscript uses "older adults without neurodegenerative pathologies" as its inclusion criterion, but this is never operationalized. How was this verified for the participants in the evaluation? Was any cognitive screening applied, or is it based on caregiver report? Given that the system is positioned partly as an alternative to clinical speech stimulation, the population characterization should be tightened.

A few smaller points the authors may want to address in passing: Figure 3 is in Spanish while the rest of the manuscript is in English, which is jarring and should be made consistent; the references to supporting materials [20], [21], [22] point to Google Drive links, which are not stable scholarly sources and will almost certainly break.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

What I am missing in the paper is that there is no new ASR method, no new pronunciation scoring method, and no comparison with alternative algorithms.

The UCD must be provided in English. Furthermore, it is not all visible. The authors must also ensure that CRUD operations are not drawn in UCD. The authors mixed a high-level use case with its internal CRUD operations. In UML, the authors should either show a single high-level use case, such as Manage Older Adult, or separate use cases, such as Create Older Adult, Edit Older Adult, and Delete Older Adult. They should not mix both levels unclearly. "interactuar con el test" is not a UC; this is clear from the association line. There are many more UCD-related problems, so a well-skilled person should redraw it. 

It is reported 90% recognition accuracy, but the recognition accuracy is based on only 50 pronunciation attempts. This is a very small sample for evaluating speech recognition, especially with older adults, where variability in accent, articulation, vocal intensity, pauses, and speech rate is expected.

I also have a small problem with the title of the paper. It comprises "speech improvement", but the study does not measure improvement in speech. 

The text says the backend coordinates integration with external Text-to-Speech services, but the core method depends on Automatic Speech Recognition. The architecture figure appears to include voice processing and external services, but the text emphasizes TTS at that point. Please elaborate. 

All in all, it looks like a nice student project with a lot of potential. It just has to be better written. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have re-read the revised manuscript alongside the authors' response letter, and I am pleased to say that the revision is thorough and responsive. The authors have engaged seriously with each of the points raised in the previous round, and the manuscript is substantially stronger as a result.