SpeakoFlow Files

SpeakoFlow 1.4.0

SpeakoFlow has its own model now. SpeakoFlow Mini is a 0.8B fine-tune that does one job, turning what you said into clean text, and it downloads at 795 MB instead of the several gigabytes a general chat model costs. This release also answers the request in issue [#18]: you can point SpeakoFlow at models already sitting on your disk instead of downloading them a second time.

SpeakoFlow Mini, a cleanup model of our own

AI cleanup used to borrow a general chat model. That works, but a chat model has to be talked out of chatting, and the small ones would narrate their plan instead of handing back your text.

Mini is trained for the job. It removes filler, fixes grammar and punctuation, and applies spoken edits, so saying "new paragraph" or "scratch that" mid-dictation does what you meant. English only for now. You can take it during first-run setup, or later in Settings, then Dictation.

Its prompt ships with it and the two are paired for you. Mini gets the short prompt it was trained on, a general model keeps the long one it needs, and a prompt you wrote yourself is never touched. Two settings that asked you to tune this by hand, cleanup strength and misheard-word repair, are gone. They were extra prose competing with the prompt for a small model's attention.

Use the models you already have

Settings, then Dictation, then Use a local model. Pick one or more .gguf or Whisper .bin files, or link a folder and every model inside it appears, subfolders included. Linked folders are rescanned when the app starts, so a model you drop in later shows up on its own.

Nothing is copied and nothing is moved. SpeakoFlow reads the file where it already sits, and removing an entry or unlinking a folder only takes it off the list. What a file is gets decided by reading the GGUF header rather than the filename, so a speech model, a chat model and a vision projector land in the right place even when all three are called model.gguf. If a file later disappears, its row says "File missing" instead of offering a download that cannot work.

Model downloads are much faster

Downloads ran at roughly 0.5 MB/s far from the origin, on a link where plain HTTP managed 10. Two causes. hyper defaults to a 64 KiB HTTP/2 flow-control window, which caps a transfer at that window divided by the round-trip time however much bandwidth you have, so downloads now speak HTTP/1.1 only. And one stream cannot fill a high-latency link by itself, measured at 10.1 MB/s on one stream against 19.1 on eight over the same connection, so the downloader now issues 8 range requests at once and writes each at its own offset.

Interrupted downloads resume where they stopped. Mini's SHA-256 is pinned, so a truncated or substituted file fails verification instead of quietly becoming your cleanup engine.

Fixes

• The assistant panel could lock clicks out of every other window on Windows. Collapsing it to the pill left Windows holding a focused window that refused to be unfocused, and a plain click on the pill could leave the app owning the mouse. Both are fixed, and dragging the pill now waits until you have actually moved it.
• Local AI ran on the integrated GPU. On a machine with both a dedicated card and an integrated one, the engine could put the model on the wrong device: measured 7.07 tok/s on the iGPU against 209.7 on the dedicated card, which is also 3x slower than the CPU. The flag the app was using, --main-gpu, does not choose where a model runs. --device does, and it is now resolved from the engine's own device list.
• Cleanup could delete a whole sentence. It shared the assistant's repetition penalty of 1.1, which is a chat setting. Cleanup is a copy task whose correct answer mostly reproduces its input, so penalising tokens already in the window penalised being right. A repetitive test transcript came back with an entire sentence missing and the remainder fused together. The penalty is off for cleanup now, and cleanup is sampled greedily rather than at the server's default temperature of 0.8.
• A failed cleanup pasted your raw text with no explanation. One event read as two unrelated bugs. The recording overlay now tells you cleanup did not run, and the log records why, usually a model file that moved or a drive that is not connected.
• Cleanup that correctly changed nothing looked broken. History now says cleanup ran and found nothing to fix.
• Text replacement rules missed capitalized triggers. A rule for "my name" did not fire on "My name", which happens at the start of an utterance, after a sentence boundary, and any time the cleanup pass punctuates your text. The same rules also matched inside longer words, so "US" to "United States" turned "USA today" into "United StatesA today". Both fixed, and substitutions surviving cleanup went from 16 of 22 to 20 of 22 across 11 dictations run twice.
• A new app icon, with a separate simplified version for small sizes. The old one was drawn from a single large image, and nine pale mint elements over a 16 pixel tray icon is under a pixel each, which is why it looked fine in the app and like a smudge in the taskbar.
• All 19 languages are complete again, including the new cleanup and local-model screens.

Known issue on macOS

After updating, macOS may still show SpeakoFlow as allowed under Accessibility while the app remains stuck waiting for access. Remove SpeakoFlow from System Settings → Privacy & Security → Accessibility, restart the Mac, then add SpeakoFlow again. See issue [#34].

Which file do I download

• Windows: SpeakoFlow_1.4.0_x64-setup.exe
• macOS: aarch64.dmg for M1 and newer, x64.dmg for Intel. macOS needs one extra step the first time you open it.
• Linux: amd64.deb or amd64.AppImage, and arm64.deb or aarch64.AppImage on ARM.

The two .app.tar.gz files are for updates, not for downloading.

Full changelog: v1.3.1...v1.4.0