Repository navigation
Replies: 4 comments 1 reply
|
Thanks, this is a great list. Here's where each one stands:
|
|
Two draft PRs are up from this thread: #568 (idea 4, spacing and capitals where a dictation meets the text before the caret) and #569 (idea 6, part 1: words from Wispr Flow's dictionary into Names and vocabulary, tested with Wispr Flow 1.6.1483). Next I'd like to do the Wispr meetings importer, next to the other importers. Wispr also keeps what it heard next to what the person changed it to, which could seed idea 1 once LokalBot has an opt-in place for corrections. Would you like that store's design as an issue first? Same question for a 'when I say X, write Y' replacements list. No rush on reviews; I'll keep at most two PRs open at a time. |
|
Numbers for idea 5, as you suggested: llama.cpp n-gram speculation on the dictation replay. (A second comment follows on the speech step, which turned out to be the bigger part of the wait.) Setup: the bundled llama-server (v1.0.0, build 12359), Qwen3.5-4B-Q4_K_M, the replay's production Compose options, an M4 Pro. The runner gained Result: A trap in the replay for speed work: it sends each dictation once per grant condition to the same server, and n-gram speculation then drafts from its own earlier answer. In conditions 2 to 5 Meeting notes, a caution: the main server is shared, so I took a first look with the notes replay ( The runner changes, the frozen set and the report are on a branch ( |
|
The speech half of idea 5. In Transcribe mode no language model is involved; after release the whole recording is transcribed again, and on a long dictation that is most of the wait. To time that exact call I added a headless replay (
On Automatic, a dictation is decoded twice. In Parakeet: its accuracy difference on this read speech is within the noise (the paired intervals include zero), though your meeting benchmarks put Qwen3-ASR slightly ahead on conversation, and Parakeet ignores the vocabulary. So this isn't a case for changing the default, but dictation and meetings may want different models. The replay, the frozen set and the report are on a branch (
|
Uh oh!
There was an error while loading. Please reload this page.
Hi! I've spent the last couple of weeks building a personal, fully local Wispr Flow replacement for my Mac (Parakeet v3 through FluidAudio, plus a small cleanup model tuned on my own dictation history). Then I found LokalBot, and it's impressive. The LiveTurns and routing benchmarks are exactly the kind of evidence I wish more projects published.
A few things from my build don't seem to have an equivalent in LokalBot yet. If any of them fit where you're taking the app, I'd be glad to write them as PRs against LokalBot's own code, following your conventions.
Learning from corrections. After a dictation is pasted, read the field (Accessibility) until the user stops editing, diff the pasted text against the edit, and learn 1–3-word swaps. Rules that kept it from learning junk:
The learned swaps run before cleanup. In my use it's the single biggest improvement over time.
A hard word lock on cleanup. Instead of relying on the prompt, align the cleanup output to the ASR words and put back any word the model added, dropped or swapped. Filler sounds (um, uh) are the only exception, and a run of words put back keeps the ASR's punctuation. Your cleanup prompt kept 16 of 16 direct sentences in your benchmark; the lock makes it a guarantee, including with smaller models.
In-person meetings. When the system-audio track carries almost no speech (I use fewer than 20 words), treat the meeting as in the room. Split the mic track by voice, skip echo removal, and let the user's own voiceprint name them.
Spacing and capitals at paste. Add a leading space only when the character before the cursor is text: use the previous paste when nothing was typed or clicked since, otherwise read the cursor through Accessibility. When the text before doesn't end a sentence, lowercase an everyday first word (never "I", names or acronyms).
Faster cleanup without a draft model. Prompt-lookup decoding, guessing the next tokens from the dictation itself and checking them in one pass, gave the same output with a median of 288 ms against 372 ms from audio to text in my tests. That's on MLX with a ~1B cleanup model; I can share the measurement method.
Import from Wispr Flow. Since LokalBot positions itself as a Wispr alternative: Wispr's local database keeps each dictation's audio, raw and cleaned text, and the user's edits. An importer could seed the dictionary, the learned corrections and a personal test set for comparing models.
Happy to open an issue for whichever you'd want first, or to share more detail here. Thanks for building this in the open!
All reactions