One pipeline, two very different tasks
The paper builds a single Bangla-language pipeline that does two jobs: it flags whether a speaker shows signs of dementia, and it predicts the next words of a partially spoken sentence. These are separate models with separate scores, and the gap between them is the most important thing to understand before reading any headline number.
The detection side works from recorded voices. The authors extracted acoustic features meant to capture vocal patterns linked to cognitive decline, then compared machine learning and deep learning classifiers on those features. They also transcribed the recordings with a transformer-based Bangla speech recognition model and ran a second set of classifiers on the text. The completion side is a separate sequential model trained to predict the next word in dementia-related Bangla sentences.
Who was recorded, and how the groups differ
The dataset contains 740 voice samples, split evenly: 370 from people with dementia and 370 without. The dementia group was clinically diagnosed and under continuous medical supervision, with Mini-Mental State Examination scores of 23 or below; the non-dementia group scored 24 or above. Participants read prepared sentences aloud rather than speaking spontaneously, a controlled setup the authors say allows reliable comparison of acoustic and linguistic features.
A critical detail sits in the demographics: the dementia group was aged 62 to 78, while the non-dementia group was aged 22 to 35. The two classes therefore differ in age as well as diagnosis. The paper does not report a test separating age effects from dementia effects, so a classifier could be picking up on voice characteristics of older speakers rather than on cognitive decline itself. Any reading of the high voice accuracy has to carry that caveat.
The text side is smaller. There are 1,890 sentences in total: 740 transcribed sentences plus 1,150 additional dementia-related sentences curated for the completion task, giving 1,520 dementia-related sentences overall. The authors themselves note that this text set is limited for models like LSTM that usually need large corpora.
Voice classification separates the groups at 97.30%
On the acoustic features, the LSTM model reached 97.30% accuracy. Its confusion matrix over 148 test samples shows 71 true positives and 72 true negatives, with 5 misclassifications. Gradient Boosting and Random Forest both reached 96.62%, and Logistic Regression trailed at 89.86% in the results table, though the text elsewhere cites 91.22% for that model — an internal inconsistency worth noting.
The authors' explanation is architectural. Tree-based models such as Random Forest and Gradient Boosting treat features as independent observations, while LSTM processes speech as a time-dependent sequence and can retain information from earlier steps. That matters if the relevant cues are things like slower speech, pauses, or repetition that unfold across an utterance rather than in a single instant. This is the paper's interpretation of why LSTM wins, not a separately tested causal claim.
The authors also emphasize the low false-negative count, arguing that missed dementia cases are the costlier error in screening because they delay intervention. That framing is reasonable for a screening tool, but it is a design priority stated by the authors, not a clinical validation result.
Transcribed text loses the acoustic cues
Text-based classification was weaker across the board. LSTM reached 82.4% test accuracy, Logistic Regression 83.7%, Support Vector Machine 81.7%, and Random Forest 72.9%. Random Forest showed clear overfitting, with 100% training accuracy against its test score, and its recall split unevenly between classes.
The paper's explanation is that transcription strips out exactly the signals that carry cognitive information: speech rate, pause duration, pitch variation, and hesitation patterns. Once those are gone, the model sees only word patterns. The authors also suggest that some misclassified dementia samples may reflect linguistic similarity between mild dementia speech and normal aging speech.
This comparison is the paper's most useful internal contrast. The same underlying recordings, routed through acoustic features versus transcribed text, produce roughly a 15-point accuracy gap in favor of voice. That supports prioritizing acoustic analysis over text-only analysis in this pipeline, while the small transcribed dataset remains a competing explanation for part of the gap.
Sentence completion works best on familiar sentences
The Bi-LSTM model with an attention mechanism, trained to complete partially expressed Bangla sentences, reached 60.75% training, 61.35% validation, and 59.74% test accuracy. Perplexity — a measure of how uncertain the model is when predicting the next word, where lower is better — was 5.27 on training, 6.13 on validation, and 6.52 on test.
The authors describe these numbers as moderate but reasonable given the task's difficulty and the limited dataset. Their manual evaluation adds a useful nuance: the model produced contextual, meaningful completions on sentences similar to its training data, but sometimes predicted meaningless or irrelevant words. That pattern is consistent with a model that has learned the training sentences more than general Bangla sentence structure.
So the completion component should be read as a feasibility demonstration. A 59.74% next-word accuracy on a small, curated set of dementia-related sentences does not establish that the system can follow a real conversation, and the authors place spontaneous speech in future work.
What the evidence supports and what it does not
The strongest claim the data supports is that acoustic features from read Bangla speech can separate a clinically diagnosed dementia group from a non-dementia group with high accuracy in this sample. The weaker claims are that text alone captures dementia comparably, or that the completion model is ready to assist patients in real time.
Three boundaries limit transfer. First, the age mismatch between groups means the voice classifier's accuracy cannot be attributed to dementia alone without further testing. Second, the reading task differs from spontaneous conversation, so performance on natural speech is untested. Third, the text and completion datasets are small, which the authors identify as a cause of overfitting and of unstable completion quality.
The paper's contribution is therefore best framed as building Bangla-language resources and showing that a multimodal detection-plus-completion pipeline is technically constructible, not as demonstrating a deployable clinical or assistive tool.
Prioritize the acoustic gap over the headline accuracy
For anyone evaluating this work, the number to weigh first is not 97.30% but the contrast behind it: acoustic models clustered near 96–97% while text models sat near 73–84%, and the completion model near 60%. That ordering tells you where the signal lives in this dataset and where the engineering is still immature.
The practical next test the paper's own design implies is a matched-age comparison, because only that would show whether the voice classifier is detecting dementia rather than the difference between speakers in their sixties and seventies and speakers in their twenties and thirties. Until then, the voice result is a promising but confounded screening signal, and the sentence completion result is the component furthest from real conversational use.
Sources used in this answer
Automatic Sentence Completion Using Bengali Audio Speech with a Transformer-Based Early Speech Recognition System
An LSTM classifier reached 97.30% accuracy on acoustic dementia detection, text-based models reached 72.9–83.7%, and a Bi-LSTM attention model completed Bangla sentences at 59.74% test accuracy.