Automatically measuring speech fluency in people with aphasia: first achievements using read-speech data
Strong clinical fluency regression method validated on noisy read speech from aphasia patients; outside core SSI modalities and use-cases.
Reading guidance
- Verdict
- full-text draft · priority medium · confidence high
- Why it matters
- Provides a solid clinical speech fluency measurement approach robust to home recording noise capable of replacing subjective aphasia fluency ratings, but not an SSI interface or silent speech decoder.
- What to trust
- Basis: full text. Coverage: high. 5 evidence records back the review.
- What is weak
- Limited to read-speech scenarios; lacks dedicated modeling for repetitions initially; uses a small hand-crafted predictor set; no spontaneous speech handling or multilingual evaluation. Limited dataset size (95 recordings from 34 participants), all reading fixed sentences; no evaluation on spontaneous speech or other languages; only three trained SLP raters as ground truth. Requires external clinical validation beyond the French read-speech protocol; limited to read speech; no spontaneous speech evaluation; uses standard PC microphones which may limit clinical integration. No silent speech decoding or interfaces; restricted to read speech only; no non-acoustic sensing; limited speaker and language diversity. Overclaim risk: medium.
- Read before
- SSI review rubric
- Read next
- SSI archive
Axes
- Task
- speech fluency regression
- Modality
- Acoustic read speech recorded via built-in PC microphones during Zoom calls.
- Hardware
- Built-in PC microphones used in Zoom videoconferencing sessions.
- Output
- labels (fluency ratings on a 5-point scale).
- Metrics
- Root-mean-square error (RMSE) as low as 0.47 with added repetition-aware predictor; Pearson correlation coefficients of 0.87 at sentence level and 0.93 to 0.96 at participant level for multiple linear regression and enhanced models.
- Evaluation mode
- Leave-one-speaker-out cross-validation predicting speech fluency ratings provided by three expert speech-language pathologists on a 5-point scale at sentence and participant levels.
- Review confidence
- high
- Overclaim risk
- medium
Expert take
The full text documents a method using forward-backward divergence segmentation on remote read-speech recordings from people with aphasia, clustering speech segments into pseudo-syllables and silent breaks to derive acoustic predictors of fluency. Multiple regression models, validated with leave-one-speaker-out cross-validation against experienced SLP fluency ratings on a 5-point scale, achieve strong correlations (participant-level Pearson r up to 0.96) with root mean squared errors below 0.5 on the scale. This demonstrates a practical, low-resource, objective, and reproducible clinical fluency scoring pipeline using common microphones despite noisy remote recording conditions. The method focuses on read speech and does not tackle spontaneous speech, more complex fluency aspects, speaker independence beyond the tested French cohort, nor does it propose a silent speech interface. Nevertheless, it provides a credible, cost-effective monitoring tool for aphasia assessment workflows, framing fluency assessment as regression over acoustic timing features. Further clinical validation and extension to spontaneous speech are necessary before wider deployment.
True value
Provides a solid clinical speech fluency measurement approach robust to home recording noise capable of replacing subjective aphasia fluency ratings, but not an SSI interface or silent speech decoder.
What changed
Canon before
Aphasia fluency scoring is usually subjective, slow, and variable across raters.
Delta from canon
Replaces manual fluency judgement with automatic regression models using engineered acoustic fluency predictors robust to noisy home recordings.
Position in field
Adjacent clinical speech analytics rather than core SSI research.
Evidence
“ techniques can be used to predict the speech fluency of PWA, as evaluated by ”
author_claim · Aims · confidence 1.00
“ All participants were recorded while reading out loud a set of sentences taken from the French version of the Boston Diagnostic Aphasia Examination. ”
validation_scope · Materials · confidence 1.00
“ The four predictors were finally combined into multivariate regression models (a multiple linear regression — MLR, and two non-linear models) to predict the average SLP ratings of speech fluency, using a leave-one-speaker-out validation scheme. ”
metric · Results · confidence 1.00
“ A forward-backward divergence segmentation and a clustering algorithm were used to compute, for each sentence, four automatic predictors of speech fluency: pseudo- syllable rate, speech ratio, rate of silent breaks, and standard deviation of pseudo-syllable length. ”
fact · Materials · confidence 1.00
“ The algorithms used in this study can constitute a cost-effective and reliable tool for the assessment of the speech fluency of patients with aphasia in read-aloud tasks. ”
deployment_claim · Discussion · confidence 0.90
Limits
Technical limits
Limited to read-speech scenarios; lacks dedicated modeling for repetitions initially; uses a small hand-crafted predictor set; no spontaneous speech handling or multilingual evaluation.
Evaluation limits
Limited dataset size (95 recordings from 34 participants), all reading fixed sentences; no evaluation on spontaneous speech or other languages; only three trained SLP raters as ground truth.
Deployment limits
Requires external clinical validation beyond the French read-speech protocol; limited to read speech; no spontaneous speech evaluation; uses standard PC microphones which may limit clinical integration.
Scope limits
No silent speech decoding or interfaces; restricted to read speech only; no non-acoustic sensing; limited speaker and language diversity.