Comment on: Parrot Speech-to-text API

The Hindi-English code-mixing capability is the genuinely hard part here. Most STT models either treat it as two separate language passes or degrade significantly at switch points mid-sentence. How is Parrot handling segmentation at the language boundary? Specifically, when a speaker switches mid-phrase rather than mid-sentence, does the model maintain a single continuous transcript or does it stitch segments, and how does that affect downstream NLU latency?