Run a real customer call from a bilingual market through an off-the-shelf transcription service and you will see the problem immediately. A speaker who moves between English and Spanish mid-sentence — as millions of people in South Florida do all day — produces transcripts full of phantom words, wrong-language guesses, and mangled names. For a media library it is embarrassing. For a medical or legal record it is dangerous.
The root cause is architectural. Acoustic models are trained to map sound to text, and they do that remarkably well — but they hold only a shallow model of meaning, domain, and context. When the audio is clean, single-language, and generic, shallow is enough. Code-switched speech, specialized vocabulary, and regional accents demand something the acoustic model does not have: understanding.
The fix is a two-stage pipeline. Stage one, a strong acoustic model produces the best raw transcript it can. Stage two, a large language model — primed with the domain vocabulary, the expected languages, and the context of the recording — corrects it: repairing code-switch boundaries, restoring technical and proper-noun accuracy, and normalizing formatting. The LLM never hears the audio; it repairs the text using knowledge the acoustic model lacks.
In our production deployments this architecture turns marginal transcripts into usable ones, and usable ones into publishable ones — at a per-hour cost that batch processing keeps trivial next to human transcription. The same design extends naturally to live events: streaming transcription with rolling correction, and translation layered on top.
If your organization serves a bilingual public and abandoned voice AI after a disappointing pilot, the technology did not fail you — the architecture did. Benchmark a two-stage pipeline against your own recordings before you write the capability off.