Abstract:
The article addresses the problem of communicative bias in automatic speech recognition systems when processing atypical speech in Russian. The relevance of the study is driven by the active integration of artificial intelligence technologies into the social sphere and public administration in Russia, which presupposes ensuring their accessibility for various categories of citizens, including people with speech disorders. The authors introduce the concept of communicative bias of ASR models as a property of a system that, under conditions of uncertainty (atypical speech, noise, pauses), makes decisions about filling semantic and acoustic gaps based on its own training data rather than the speaker's actual communicative intention. The empirical basis of the study is the RuAphasiaBank corpus, which includes speech recordings from patients with various types of aphasia of varying severity, as well as neurotypical informants. For comparative analysis, six ASR models operating with Russian were selected: two open-source (GigaAM, T-One) and four commercial (Yandex SpeechKit, Any2Text, Charla, Speech2Text). Recognition quality was assessed using WER and CER metrics, as well as a qualitative analysis of the ratio of insertions, deletions, and substitutions at the word and character levels. The results show that the models implement different strategies when encountering atypical speech: “panic shortening” (GigaAM, Yandex SpeechKit), “lexical preservation” at the expense of accuracy (T-One), and a “stable” strategy of balanced errors (Any2Text, Speech2Text, Charla). Linguistic analysis reveals that each model produces a specific distorted communicative style, ranging from colloquial dialogic speech to a style associated with cognitive impairments or the use of youth slang. Based on the obtained data, potential risks to the social well-being and legal security of aphasia patients when using voice assistants in public institutions are modeled.