Episode Details

Most people assume Whisper is the default for speech-to-text everywhere. But Whisper was built for server batch transcription, not real-time phone dictation. NVIDIA Parakeet uses a radically different architecture — CTC instead of autoregressive decoding — that lets it output tokens every 80 milliseconds as you speak. This episode breaks down why Parakeet runs faster on a phone chip than Whisper Tiny despite having 600 million parameters to Whisper's 39 million, where the accuracy tradeoffs actually bite, and how punctuation handling differs between the two approaches. If you're building an AI keyboard or dictation feature, this comparison will change how you think about model selection. Episode #629362 — open it directly at myweirdprompts.com/629362