ServiceNow and Hugging Face released benchmarks testing how frontier ASR (automatic speech recognition) models handle code-switched speech—bilingual or multilingual utterances common in global customer support. Frontier models showed wide variance in handling language mixing, with top performers at 85%+ WER on code-switched Spanglish.
The findings matter for voice agents deployed in multilingual contact centers and IVR systems. Developers targeting emerging markets or immigrant-dense regions now have baseline data to audit model behavior before production deployment, improving accessibility for non-English-monolinguals.