A user-centered design approach guided the project from early requirements gathering to final prototype validation [23,24]. Initial requirements were collected before the frontend design to match the SLPs’ requirements for a web-based digitized speech and language assessment [22]. This study consisted of a three-step approach: (1) needs and context analysis including specifying the context of use and gathering user requirements; (2) preliminary testing; and (3) end-user testing and refinement. The 2 empirical components—prototype development and preliminary testing, and end-user testing and refinement—are reported in the “Results section.” Two iterative usability testing cycles were conducted to identify and address initial usability issues and to evaluate the effectiveness of implemented improvements in a refined prototype.

Patient data were obtained from patients with stroke and the publicly available TORGO database, a database that includes individuals with cerebral palsy and amyotrophic lateral sclerosis, which permits academic use [25]. All procedures adhered to ethical guidelines and regulations. Approvals were obtained from the Ethics Committee of Northwestern and Central Switzerland (Req-2024‐00103 for usability testing, R-2025‐00538 for voice recordings). Written informed consent was obtained from all participants. The participants did not receive compensation.

The prototype was evaluated in 2 stages: prototype development and preliminary testing with lay users to detect usability issues and bugs, as well as the technical validity of ASR-based intelligibility scoring; and end-user testing and refinement with SLPs across 2 iterative cycles, during which SLPs evaluated usability, workflow integration, and clinical relevance.

Local community members familiar with laptops explored the prototype freely, uploading recordings and navigating the dashboard. A mock session ensured comparability. Minor issues (eg, broken links and unclear navigation) were corrected before expert testing.

Eight lay users (Table 1) transcribed audio recordings from 3 patients with dysarthria to validate the accuracy of the automatic intelligibility assessment. The recordings included the FDA-2 words and sentences tasks. Performance was quantified using WER, with lower values indicating greater accuracy.

The ASR system (Whisper Large-v3) performed comparably to, and in some cases better than, untrained listeners. For mild dysarthria, ASR achieved lower median WERs than most participants. For moderate and severe dysarthria, ASR errors increased but remained within the variability range of human listeners. Boxplots of WER distributions are shown in Figure 5.

These findings confirm that ASR can approximate human perceptual judgments of intelligibility and support its integration into the prototype.

Eight SLPs participated in the first usability cycle (mean clinical experience 7.56, SD 3.2 y; Table 1). Most accessed the application at the workplace, while 2 participated from home, using a mix of Chrome, Firefox, and Edge.

The mean SUS score was 88.4 (SD 4.6), which is considered excellent [38,39]. Average ratings across survey items were 4.26/5 (Table 2). Core workflow elements—including navigation, uploading sessions, segment selection, and processing time—scored between 4.50 and 4.63. Visual design received the highest rating (mean 4.88, SD 0.35). Understanding the intelligibility outputs scored slightly lower: the side-by-side transcription view was rated 4.50, and the average intelligibility score clarity was 4.13 (SD 0.13). The lowest ratings were for the perceived accuracy of the automatic intelligibility score (mean 3.13, SD 1.25) and phoneme-level error highlighting (mean 3.88, SD 1.25). Qualitative feedback (Table 3; Table S9 in Multimedia Appendix 1) clustered into 3 categories:

These results confirmed high usability while identifying areas for refinement before the second cycle.

Five SLPs participated in the second usability cycle (mean clinical experience 4.3, SD 2.0 y; Table 1). The testing environments and browsers were the same as those used in cycle 1.

The mean SUS score increased to 91.7 (SD 4.1), which is again considered excellent [38]. Survey results (Table 4) showed consistently high ratings. Navigation and the new acoustic analysis page both received a score of 4.75, indicating that participants found the workflow intuitive and the additional analysis features useful. Frequency and intensity graphs were rated positively (4.33 each), while the formants graph was rated lower (mean 3.66, SD 0.65), suggesting limited clarity or perceived clinical utility. Overall expectations were rated at a mean of 4.20 (SD 0.42).

Representative feedback (Table 5) emphasized advanced feature requests. Suggestions included the option to add notes to assessment results, integration of advanced voice-quality indices (eg, Cepstral Peak Prominence [CPPS] and Acoustic Voice Quality Index [AVQI]), and audio playback directly from result pages. These comments indicate readiness for clinical refinement, with priorities shifting toward advanced analysis rather than fundamental usability.

Across all 3 substudies, the iSpeak prototype demonstrated promising performance and usability. The pilot validation confirmed that ASR achieved accuracy comparable to or greater than that of untrained human listeners across varying severities of dysarthria. Both usability cycles with SLPs yielded excellent SUS scores (>88), with cycle 2 reaching a mean of 91.7 (SD 4.1) after refinement. Core workflows were consistently rated highly, and feedback evolved from identifying technical issues in cycle 1 to requesting advanced analytical features in cycle 2. Together, these results indicate that the system is both usable and potentially clinically relevant, with future development focused on enhanced analytic capabilities and broader validation.

This study evaluated a web-based application designed to support SLPs in assessing dysarthric speech in patients with poststroke. Developed using a user-centered design framework in accordance with the International Organization for Standardization (ISO) 9241‐210, the prototype integrates ASR and acoustic analysis into a clinical dashboard. A pilot validation compared ASR with lay listeners, followed by 2 iterative usability cycles with SLPs. The findings demonstrate that ASR can approximate human performance in intelligibility assessment and that the application was consistently rated as highly usable, with SUS scores above 85 in both cycles. Feedback confirmed the value of automatic intelligibility scoring and acoustic analysis while identifying priorities for further development.

The pilot validation provided initial evidence of how ASR compares with human listeners on dysarthric speech. Across tasks, word lists produced higher WERs than sentence lists, reflecting the contextual advantage of sentences [27]. Importantly, lay listeners were not native English speakers, which may have introduced bias, as subtle dysarthric articulations could be more difficult to interpret for nonnative listeners [40]. This may have led to an underestimation of intelligibility relative to trained or native listeners [41].

Patient-level differences were evident. For “Patient 001,” ASR outperformed most lay listeners across tasks, suggesting that the system could handle this speech relatively well. For “Patient 003,” ASR performance was competitive, especially in the sentences task. “Patient 002” posed the greatest challenge: lay listeners showed high variability in WERs for the word task, and ASR performed worse than all listeners in the sentences task. These findings highlight how severity and individual variability in dysarthria strongly affect ASR performance [27,40,41].

Although Whisper achieves state-of-the-art performance on typical speech, its accuracy declines with dysarthric input. Stroke-related dysarthria alters pronunciation, speech rate, and voice quality in diverse ways. Because such data are underrepresented in training corpora, ASR robustness remains limited [32,41,42]. Nonetheless, pilot results indicate that Whisper can sometimes match or surpass untrained human listeners, suggesting promise as a baseline for automatic intelligibility scoring. Similar findings are reported in recent validation efforts of automatic intelligibility measures for motor speech disorders [43]. Future validation with larger patient cohorts and SLP raters is needed to confirm these observations.

The first usability cycle demonstrated that the prototype was already perceived as highly usable. The mean SUS score of 88.4 (SD 4.6) falls in the “excellent” range, confirming that the design was well-adapted to SLPs’ needs [44]. Importantly, the application was tested both in clinical and home environments, as well as across different browsers, indicating technical robustness in varied contexts of use.

Survey ratings averaged 4.26/5. Core workflow elements—navigation, uploading, segment selection, and processing time—scored above 4.5, and visual design was rated highest (mean 4.88, SD 0.35). These results align with reviews of existing speech therapy apps, which similarly show that visual design and general usability are often rated highly while output clarity (such as the accuracy of automated scores or error highlighting) tends to receive more mixed feedback [38,45].

Three items scored lower: perceived accuracy of the intelligibility score (mean 3.13, SD 1.25), phoneme-level error highlighting (mean 3.88, SD 1.25), and meeting expectations (mean 3.88, SD 1.13). These reflected a technical alignment bug in this prototype version, which caused intelligibility scores and error highlights to miss clearly spoken words. This limitation directly influenced perceptions of accuracy, a pattern also reported in empirical studies of online SLP tools, where users frequently note misalignments or ambiguous output explanations [39].

Open feedback aligned with these quantitative findings. Fixes focused on reliability (eg, failed uploads and region shifting). Usability improvements included clearer terminology, simultaneous listening and transcription, and visible time markers. Feature requests suggested phonetic equivalence in scoring and pause analysis. Similar issues of clinician-perceived usability and workflow integration have been noted in broader reviews of automated speech therapy tools [45,46]. Fourteen priority fixes and improvements were implemented for cycle 2, balancing feasibility with the need to proceed rapidly to the next evaluation.

The second cycle reinforced and extended these findings. The mean SUS score increased to 91.7 (SD 4.1), again in the “excellent” range, suggesting that refinements made after cycle 1 successfully improved usability, as SLP tools evolve from basic usability to richer features [44,45]. As before, testing across home and clinical settings and multiple browsers confirmed robustness. This pattern mirrors findings in reviews of eHealth speech-language therapy applications, where iterative improvements based on clinician feedback tend to yield measurable increases in satisfaction and functionality [38,45].

Survey ratings averaged 4.34/5. Navigation and the new acoustic analysis page received the highest scores (mean 4.75, SD 0.78). Frequency and intensity graphs were positively rated (4.33 each), confirming clinical relevance. The formants graph was rated lower (mean 3.66, SD 0.65), indicating that its value was less evident in practice. The “meeting expectations” item improved to a mean of 4.20 (SD 0.42). These improvements in visual clarity and interactive analytics reflect the user experience findings from telepractice tools, where dashboards and visualization features are increasingly valued as maturity grows [47].

Qualitative feedback shifted from identifying bugs to requesting advanced features. Suggestions included adding notes to assessment results, incorporating voice-quality indices such as CPPS and AVQI, and enabling audio playback directly from results pages. This shift is consistent with empirical studies of speech therapy platforms, where, once baseline usability is achieved, user requests focus more on analytic depth and interactivity [20,48]. Broader surveys of online and AI-enabled speech therapy systems echo this pattern, with early iterations focusing on usability and later development emphasizing advanced functionality [45,46].

The final prototype can be considered a simple, efficient, and potentially clinically relevant dashboard for assessing speech clarity and phonetic accuracy. Usage analytics confirmed the intuitive design: task times decreased after initial use, and only 3 dead clicks were recorded. The modular architecture enables extension to multilingual contexts, although current phoneme segmentation is restricted to English. While Whisper is multilingual, the phoneme module relies on English-only resources, limiting its applicability across languages.

Occasional ASR hallucinations were observed in cases of severe dysarthria or poor microphone quality. These issues were mitigated through design decisions (eg, suppressing insertions in alignment), ensuring that phoneme analysis was not distorted. The current implementation also assumes one speaker per session; while speaker diarization could be added to handle overlapping speech, it introduces additional complexity.

Several directions for development emerged. First, multilingual phoneme support is required to reflect the diversity of clinical populations. Second, advanced acoustic indices (eg, jitter, shimmer, CPPS, and AVQI) should be integrated to capture additional aspects of dysarthria severity [35-37]. Third, diarization would allow for the analysis of multispeaker sessions, accounting for overlaps or caregiver contributions. Finally, larger-scale validation with patients and practicing SLPs is essential to confirm accuracy and usability across broader contexts. An important next step is to systematically evaluate the generalizability of the system across different neurological populations, as initial testing across mixed etiologies suggests feasibility but does not yet establish condition-specific validity. Future work should also consider implementation within the broader framework of telerehabilitation, where evidence is growing for the effectiveness of remote interventions in speech-language therapy [49].

At present, the system is not compliant with data protection regulations such as the Health Insurance Portability and Accountability Act (HIPAA) and the Health Information Technology for Economic and Clinical Health Act (HITECH), which poses a barrier to clinical deployment over the internet. Use within closed clinical networks is feasible, but full compliance will be essential for wider adoption. A limitation of this study is the small sample size in the usability testing cycles, which may limit the generalizability of the usability findings and warrants further testing in larger cohorts.

This study demonstrates the feasibility of integrating ASR and acoustic analysis into a web-based application to support SLPs in assessing dysarthric speech. A pilot validation confirmed that ASR performance was comparable to that of untrained human listeners, while 2 usability cycles with SLPs yielded consistently excellent SUS scores, indicating that the system was perceived as highly usable and clinically relevant. Iterative refinements improved navigation and workflow, and feedback evolved from bug reports to requests for advanced analytic features, underscoring both the robustness of the core design and the demand for deeper functionality. Although current limitations include reliance on English-only phoneme segmentation, the prototype establishes a solid foundation for scalable digital assessment. Future work should extend validation to larger and more diverse patient groups, expand multilingual support, and integrate additional advanced outcome measures to further enhance clinical adoption and impact.