Count the GPU call from outside, where the result survives

Keep our own books on the GPU budget

Report how fast the run actually was

Fewer GPU calls, faster attention, optional second pass

Do not carry an expiring token into a long job

Actually set the estimate before the first section

Estimate remaining time from measured pace

Convert the audio in the job, not in the request

Save the background result instead of holding it in RAM

Hand pywebpush the key in a format it accepts

Restore the lost binding in the push sender

Give the push send a deadline and a valid sub claim

Make push verifiable instead of silently dead

Carry the caller's context into the worker thread

Run the pipeline server-side and push when done

Probe: which quota does background work draw on?

Probe: can GPU work start without a live request?

Cross-check, confidence, voice prints, quality, learning

Cut in silence, drop invented subtitle text

Make the glossary correct instead of decorate

Give Whisper its temperature fallback

Report newly learned glossary terms

Split pure logic into kern.py; delete audio after analysis

Remove the language-model smoothing pass

Never rewrite a transcript the user supplied

Keep the raw transcript and mark heavily smoothed lines

Drop the Whisper prompt that silenced recognition

Automatic ranked glossary and language-model smoothing pass

Persistent glossary feeding recognition across recordings

Make section processing idempotent for safe retries

Do not depend on model timestamps; align window text instead

Use a German-finetuned Whisper for German audio

Word-level speaker assignment, audio normalisation, beam search

Add status endpoint reporting HF token presence

Optional: align a supplied transcript instead of running ASR

whisper-large-v3 and pinned language to stop English translation

Stepwise API so long recordings survive GPU token expiry

Fix window boundary rounding past end of file

Support long recordings: windowed transcription, progress

Size ZeroGPU reservation by audio length

Full-audio transcription, language detection, save with speaker names

Include generated markdown in analysis response

Transcode uploads to WAV before analysis

Turn Space into key-protected JSON API backend

Polish history cards and hide empty file box

Mobile-first UI: responsive CSS, card history, touch targets

Single-flow UI: full analysis without use-case selection

Disable SSR mode (incompatible with Gradio auth on Spaces)

Add app-level login (credentials from secrets)

Enable PWA mode