One Model or Two? ASR Cleanup Explained

Whisper bakes punctuation into its weights. Parakeet-TDT emits it natively from a transducer. And yet a 484-megabyte text model called S1-mini exists purely to clean up after speech recognizers — and it never hears a single second of audio. So where does recognition actually end and cleanup begin? This episode works through the architecture question from the ground up: CTC, RNN-T, TDT, encoder-decoder transformers, and the newest audio-language foundation models, and what each one can and cannot emit. The answer turns out to hinge less on topology than on training targets — and on a metric that didn't exist until recently. Key sources: • Superwhisper blog, Introducing the S1 family of models, 2026-08-19 — https://superwhisper.com/blog/s1 • Hugging Face, superwhisper/s1-mini model card (release v1) — https://huggingface.co/superwhisper/s1-mini • arXiv:2212.04356, Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (Whisper) — https://arxiv.org/abs/2212.04356 Episode 5974 — myweirdprompts.com/n/5974 (permanent short link: myweirdprompts.com/558256)