• •Hugging Face researchers evaluated current Automatic Speech Recognition (ASR) systems on bilingual code-switched audio samples.
• •Models demonstrated high accuracy when speakers stayed within a single language but saw performance drops when switching languages mid-sentence.
• •It remains unclear whether these systems can be fine-tuned to maintain accuracy without sacrificing general language proficiency.

A recent analysis from Hugging Face highlights that current voice agent ASR systems often fail to process speech when users frequently switch between two languages in a single conversation. While models perform reliably during monolingual exchanges, performance degrades significantly during code-switching, a common trait in multilingual communities. However, researchers noted that the datasets used for training often lack sufficient examples of such fluid language transitions, leading to unpredictable errors. Whether these limitations can be corrected through better training data or architectural changes remains a subject of ongoing investigation.