Parakeet TDT v3 still leads CPU dictation tradeoff

This is a classic case where leaderboard winners and user-facing winners are not the same thing.

• –Hugging Face’s Open ASR leaderboard puts Canary-Qwen 2.5B ahead on WER, but the thread centers on the much bigger latency penalty for CPU-first local dictation
• –NVIDIA’s Parakeet TDT 0.6B v3 model card positions it as a high-throughput multilingual ASR model, which matches why developers keep reaching for it in offline transcription apps
• –For hold-to-talk dictation on Windows, perceived responsiveness matters more than absolute benchmark rank, so a slightly weaker model can still be the better product choice
• –The interesting gap is not accuracy alone but deployment profile: Parakeet is being used via ONNX on CPU, while stronger rivals are often treated as GPU-class models
• –This makes Parakeet less a “best overall ASR model” story than a “best local inference compromise” story for practical desktop tooling