Automatic classification of aphasia severity presents persistent challenges, particularly for languages with limited clinical speech resources such as Russian. This paper explores a multimodal approach to severity estimation that combines acoustic and semantic representations of pathological speech. Acoustic features are extracted using pretrained Wav2Vec 2.0 models, while semantic information is obtained from the encoder of the Whisper model. The two representations are integrated via early feature fusion and evaluated using gradient boosting classifiers in a speaker-independent cross-validation setting. Experiments are conducted on a newly collected dataset of Russian speech recordings from patients with aphasia and neurotypical speakers (RuAphasiaBank). The results suggest that the combined use of acoustic and semantic embeddings can provide more stable severity estimates than unimodal baselines. This study contributes empirical evidence on the applicability of multimodal representation learning for aphasia severity classification under data-scarce conditions.

Texts are widely used in aphasia rehabilitation to support the recovery of comprehension and narrative planning. In routine practice, clinical impact depends strongly on patient motivation and on the availability of age-appropriate reading materials: adults are often offered child-oriented texts, which can be perceived as demeaning and may reduce engagement. We present a controllable generation pipeline for building a repository of Russian therapeutic reading texts for adult aphasia therapy. An anonymized repository with code and data is available at https://github.com/z00logist/aphasia-exercises-generation. The pipeline conditions each story on an explicit semantic triplet (12 topics-10 archetypes-11 objects) and enforces three clinically motivated complexity regimes (Basic/Intermediate/Advanced). Using batched prompting, we generate 1,296 unique stories. We evaluate the corpus with classical linguistic metrics and a LLM-as-a-judge protocol (18 binary criteria); on a stratified sample of 198 stories, overall rubric compliance is 80.0%. Surface metrics show a monotonic increase in lexical and syntactic complexity across regimes, and Basic texts closely match a small clinical anchor set of 10 therapist-authored texts. Judge-based analysis indicates near-perfect adherence to high-level narrative constraints but persistent limitations in fine-grained phonotactic control, motivating hybrid neuro-symbolic enforcement.