Can Voice Agents Handle Bilingual Customers? Benchmarking Frontier ASR on Code-Switched Speech

The more pointed finding here is not simply that frontier ASR degrades on code-switched speech, but that the degradation is measurable enough to matter at production scale, meaning the gap is not a theoretical edge case but a quantifiable reliability problem that procurement teams and voice platform builders can now cite with numbers.

This is largely disconnected from recent activity in our archive, as Modelwire has not yet covered ASR benchmarking or multilingual voice infrastructure. The story belongs to a broader conversation happening across the speech and voice agent space, where deployment teams are discovering that models optimized on clean, monolingual corpora perform poorly against the messy reality of how bilingual populations actually speak. That gap between benchmark conditions and production conditions is a recurring theme in applied ML, and this paper gives it a concrete, reproducible form in the ASR domain.

Watch whether Whisper or any other major ASR provider publishes a targeted code-switching fine-tune or training data release within the next two quarters. If they do, this benchmark will have functioned as the forcing document that moved the roadmap.