www: put the two engines and the 4-bit options on the site

[z33b0t]

The site still described a Parakeet-only, bf16-only server. Its last content change was 907ac19; every www/ commit since is dependabot. Four model features shipped in that window and none of them appear anywhere:

shipped	the site said
fp16 checkpoint (#68)	"any GPU from the 20-series up" — accidentally made true by it, never explained
nvfp4 for Parakeet (#69)	~2.3 GB VRAM
Whisper engine (#71)	no mention at all; 25 languages; pipeline hard-coded Parakeet
nf4 for Whisper (#72)	—

The two most valuable things we can now say — ~100 languages and a model that fits in half a gigabyte — were both missing.

What changed

• New "Two models. Pick one." section — Parakeet (default) and Whisper (--engine whisper) side by side, with languages, WER, latency, and VRAM for each, plus a note that both take a 4-bit option and why the formats differ.
• The naming joke becomes the reveal. The page's whole gag is that the field is "the word Whisper with the vowels moved around." Rather than fight that, it now lands as: we didn't need to name ourselves after Whisper — we just run it when you want it.
• Pipeline diagram no longer hard-codes Parakeet.
• Stats, card 02, and the systemd block re-numbered.

Numbers were re-measured, and two were wrong

Everything is measured on an RTX 5090 over the repo's own 208-clip / 25.8-min corpus, not carried over:

• ~70ms to transcribe a 35-second clipconflated two figures. 75 ms is the median dictation; a 35.6 s clip takes ~277 ms. The stat is now labelled for what it measures.
• ~2.3 GB VRAM → 1.4 GB peak in bf16, 0.5 GB in 4-bit.

Parakeet (measured, isolated)	weights	peak	median clip	35.6 s clip
bf16	1.28 GB	1.43 GB	74.9 ms	276.6 ms
nvfp4	0.51 GB	0.66 GB	176.4 ms	428.5 ms

A bug this turned up in scripts/compare_precision.py

Worth recording separately: run in one process, that script's VRAM figures are contaminated. Measuring bf16 and nvfp4 together, it reports Parakeet nvfp4 at 1.82 GB of weights — more than bf16's 1.28 GB, and bf16 peak at 3.84 GB against the committed table's 1.43 GB. 1.28 + 0.51 ≈ 1.82 gives it away: NeMo keeps model instances alive across the harness's reloads, so the previous precision is still resident when the next is measured.

Measured one precision per process, the numbers match the committed README table exactly (bf16 peak 1.43 GB, nvfp4 weights 0.51 GB), which is what this PR uses. The WER figures are unaffected and reproduce (2.99% / 2.86%, p=0.69). compare_whisper_quant.py in #72 already isolates per variant for this reason; compare_precision.py has not been fixed here — happy to do that separately.

Notes

• Depends on asr: add a 4-bit nf4 option for the Whisper engine #72 for the nf4 numbers to be true of main; land that first.
• Layout not visually verified — the local preview server was reachable by curl but Chrome showed an error page for it, so I stopped rather than keep poking. next build, TypeScript, and static generation all pass, and the new section reuses the card/grid classes the existing "why bother" section already uses.

🤖 Generated with Claude Code

https://claude.ai/code/session_01376qJiterkkQwTWmN1AC4Q