Question about the old Fast dictation mode

[postoso]

Hey, I had a question about the old Fast dictation mode / Parakeet fast finalization path.

Since updating to 1.6.2, I’ve noticed dictation takes longer to finish after I stop recording. It is not a huge deal for short dictations, but I often use FluidVoice for longer brain dumps, normally 5-20 minutes and occasionally 30+ minutes.

I checked the logs because I was curious what changed. It does not look like text insertion is the slow part. In my logs, insertion is usually around 0-1ms once the final text is ready. The delay seems to be before that, where 1.6.2 is doing the full final transcription pass before returning text.

My girlfriend has also been using FluidVoice for a while and noticed the same thing. I grabbed her logs too, and they showed a similar pattern.

I did a quick A/B with the same ~82s audio:

• 1.6.1 with ParakeetFinalizationMode=tokenTimedChunkMerge: source=livePreview, stop_end totalMs=699, insertion 1ms
• 1.6.2 current: source=full, provider_final_done elapsedMs=1565, stop_end totalMs=1659, insertion 1ms

And in a real 1.6.2 log from one of my longer dictations, a ~5.3 minute recording had provider_final_done elapsedMs=4979 with source=full.

I can see why full finalization might be the better default, especially if someone is inserting directly into emails, documents, forms, etc. But a lot of my usage is more like raw brain-dump input into Claude/Codex. In that workflow, slightly imperfect ASR is usually fine because the downstream model can recover the intent, but waiting at the end of a long dictation is pretty noticeable.

So I was curious: was the old Fast mode removed mainly because of accuracy/reliability issues? And are the new faster recording-start changes actually in conflict with the old fast finalization path, or could they coexist?

If it is mostly a reliability/defaults concern, would you be open to an opt-in Fast mode coming back for long-form or AI-bound dictation? I’d be happy to test or help with a PR if that direction makes sense.

[altic-dev]

how long does it take?

Yeah, the fast mode was causing a little confusion and was causing delays for shorter dictation, so I removed it from experimentation, Actually could be useful for longer sesions, hmm. How long does it tkae now? Long form makes sense actually. but idk how to not make people complain after enabling it. it's a hard problem to satisfy everyone ...

[postoso]

for longer dictations it can easily take like 20-35 seconds. i havent recorded, i can try next time. it's happens enough that sometimes i get worried it bugged and and that it wont actually work. i know the same is true of my gf. Yea it is hard because the faster dictation is nice too and i imagine most people use it for shorter bursts, altho i bet with AI a lot more people are doing like 20+ minute rambles. what was the confusion and how long were the delays for shorter dictation? i can try to think through an implementation that would help both sides of the issue if you want, just wanted to check in before spending time trying to work on this problem