Introduction
Clinical interviews are essential for neurological assessment, yet the manual transcription and analysis of these dialogues represent a significant operational bottleneck. Automatic speech processing offers a potential solution, but clinical speech—particularly from patients with neurodegenerative conditions like Huntington’s Disease (HD)—presents unique challenges. These patients often exhibit distinct prosodic changes and speech impediments that standard, off-the-shelf algorithms are not designed to handle.
The core technical challenge is speaker diarization: determining "who spoke when." This study evaluates whether it is more effective to treat this as a classification problem (Speaker Role Recognition) or an identification problem (Speaker Enrollment) within the specific context of clinical interviews.
Methodology
The researchers compared two distinct architectural philosophies to address the diarization challenge in a two-speaker (clinician and patient) environment:
- Speaker Role Recognition (SRR): A "top-down" approach that leverages the known structure of the interview. Since the roles are fixed (Neuropsychologist and Interviewee), a single multi-label neural network is trained to detect both roles simultaneously. This model utilizes SincNet filters to process raw audio waveforms, followed by LSTM layers.
- Speaker Enrollment (SE): A traditional "bottom-up" modular pipeline. This involves Voice Activity Detection (VAD), Speaker Change Detection (SCD), and an Enrollment phase where a voice "template" (embedding) is created from a known snippet of each speaker's voice. New segments are then matched to these templates using cosine distance.
Both architectures were tested using "Pretrained" models (trained on general datasets like VoxCeleb) and "Retrained" models (fine-tuned on specific clinical data).
Experimental Setup
The study utilized a specialized dataset of 94 conversational interviews conducted in French. The participant pool included 22 healthy controls, 18 "pre-manifest" HD carriers, and 54 patients with manifest HD. To ensure rigorous evaluation, the data was split into meta-train (60%), meta-dev (20%), and meta-test (20%) sets, with ground truth labels provided by graduate students in speech pathology.
Key Results
The primary metric for evaluation was the Identification Error Rate (IER). The results demonstrated a clear advantage for the end-to-end approach:
- SRR Superiority: The SRR approach achieved the best performance with an IER of 19.5%. By handling the entire process in a single step, it avoids the "error propagation" inherent in modular pipelines where a failure in VAD or SCD compromises all subsequent steps.
- The Importance of In-Domain Data: Models pretrained on general speech performed poorly. Retraining on actual clinical dialogue was essential for achieving usable accuracy. As shown in the results, the SRR model requires approximately 28 interviews (50% of the training set) to begin outperforming the SE pipeline.
- The Enrollment Bottleneck: When the SE model was provided with perfect, human-annotated segments (Ground Truth Segmentation), its error rate dropped to 9.1%. This indicates that the primary weakness in the enrollment protocol is not the voice matching itself, but the initial detection of speaker boundaries.
Clinical Implications and Robustness
Notably, the performance of these models remained robust across different patient groups. Accuracy did not significantly degrade when processing speech from patients with severe symptoms compared to healthy controls. This suggests the models are successfully identifying structural role cues, such as turn-taking patterns, rather than relying solely on acoustic clarity.
Furthermore, the SRR model proved effective at predicting the Ratio of Silence, a critical clinical marker used to track the progression of Huntington’s Disease. While other markers like the standard deviation of utterance duration were more difficult to capture, the success with silence ratios demonstrates the potential for these tools to assist in longitudinal clinical monitoring.
Conclusion
The study concludes that an end-to-end Speaker Role Recognition approach significantly outperforms traditional modular enrollment protocols for clinical applications. By reducing error propagation and leveraging in-domain retraining, SRR provides a more reliable path toward automating the analysis of clinical dialogues. Future work will focus on scaling these models to multi-speaker environments and integrating a wider array of clinical biomarkers directly into the model's learning objectives.
Summary updated