Episode Details

The Hugging Face speech recognition catalog looks like thousands of competing models, but most of it is fine-tunes, quants, and format ports sitting downstream of a handful of genuine architectural bets. This episode tours the mechanisms behind four of them — Moonshine's padding-free encoder and its ergodic streaming variant, Zipformer's U-Net-like multi-scale design with BiasNorm and ScaledAdam, SenseVoice's multi-task head for emotion and audio events, and WeNet's two-pass streaming unification — and then looks at what the derivative layer means for anyone choosing a base model. The proof case: search Parakeet on the Hub and the top result is a community MLX port out-downloading NVIDIA's original four to one. If you're picking a base model for something like English–Hebrew code-switching, the question isn't which model has the best word error rate. It's which axis the model was actually built to optimize. Key sources: • Moonshine, arXiv, revised 2024-10-22 — https://arxiv.org/abs/2410.15608v2 • Moonshine v2, arXiv, 2026-02-12 — https://arxiv.org/abs/2602.12241v1 • HF model card, created 2026-01-06 — https://huggingface.co/moonshine-ai/moonshine-streaming-tiny Episode 6052 — myweirdprompts.com/n/6052 (permanent short link: myweirdprompts.com/457719)