Automatic keyphrase extraction, once dominated by statistical tricks, is now leaning on large language models (LLMs) and full speech pipelines, a trend highlighted in the latest survey of the field. The review shows how recent work unites text‑ and speech‑based methods into a single framework, noting the shift away from simple frequency‑based scoring toward neural models that can capture deeper semantics.

In Russian speech, researchers have started to probe whether sound itself can signal keyphrases. Using the OpenSMILE toolkit and the eGeMAPS‑v0.2 feature set, one study examined manually‑identified keyphrases in popular‑science talks. The analysis found that several acoustic cues—smooth drops in pitch, pauses, and changes in spectral energy—are linked with the spoken mentions of important concepts.

However, pure statistical approaches still struggle with languages that have complex morphology, such as Russian. Because a single lexeme can appear in many inflected forms that each occurs only rarely, frequency‑based methods often miss vital phrases. This limitation explains why many existing extraction algorithms, originally tuned for English, need adaptation when applied to Russian or other morphologically rich languages.

Together these points illustrate a field in transition: researchers are combining neural language understanding with prosodic analysis, yet must also grapple with linguistic diversity that challenges traditional statistics.

[Visual: Illustration of acoustic features linked to keyphrases in Russian speech, taken from the cited survey]