Apple researchers presented a method for detecting language changes within an utterance using language-specific responses inside a speech model. Using only about 1% of the dimensions in one Whisper-small encoder layer, it achieved a mean F1 of 0.798 on CS-FLEURS, pairing English with 14 languages, compared with baselines of 0.776 and 0.751. It scored 0.734 on Chinese-English conversational data. The intended role is filtering candidates before transcription or large-language-model analysis, but actual uncurated audio-collection pipelines have not yet been tested. This is not an announcement of a released Siri feature.