Vexa weighs Parakeet, Voxtral for live transcripts

This is the kind of infrastructure question that matters more than leaderboard wins: Vexa is testing where speech models break in real deployments, not just which one tops a benchmark. It also shows how quickly teams shipping Whisper into production are running into edge cases that push them toward streaming-first ASR alternatives.

• –Vexa already runs sub-second transcript streaming over WebSockets, so the evaluation is about end-to-end behavior under production constraints rather than toy demos
• –NVIDIA positions Parakeet v2 as a high-speed, high-accuracy ASR model, but its English-first framing leaves multilingual coverage as a real risk for Vexa’s Croatian, Latvian, Finnish, and French users
• –Mistral markets Voxtral as outperforming Whisper large-v3 on speech tasks, but Vexa is explicitly asking for latency, memory, and failure-mode data that vendor benchmarks rarely show
• –The silence-hallucination angle is the sharpest part of the post: if CTC/transducer models really avoid Whisper’s dead-air failure mode, that is a meaningful operational win for live meeting products
• –Because Vexa supports both self-hosters on consumer GPUs and larger cluster deployments, model size alone is not enough; runtime characteristics, batching behavior, and degradation under load will decide what actually ships