You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
OpenAI-compatible services accept the standard OpenAI audio format. The chat-style endpoint
used to send audio only the way Alibaba Cloud Bailian expects it, so gateways that follow the
standard OpenAI format (Gemini, for example) answered with a 400. Both formats now work: auto
detection tries them in turn, or you can pick one under "Protocol". Existing configurations keep
their current request format.
AI cleanup supports OpenRouter. It can share one key with speech recognition. The default
model is DeepSeek V4.1 Flash, and leaving the address empty uses the official endpoint. In the
English UI, new AI services default to OpenRouter.
Selection edits can use a separate AI service. Pick one under AI cleanup → Context awareness,
for example a stronger model for translating or rewriting; regular dictation keeps using the
current AI service. In server mode, selection edits that use a separate service are sent from
this computer.
Changed
Context awareness has moved to the AI cleanup page. It only works while AI cleanup is on, so
it now sits next to that switch instead of in the general settings.
Tidier settings. Extra explanations on pages and cards now sit behind the ⓘ next to each
title; switches are always on the right; cleanup presets show a one-line description instead of
the start of their prompt; "Per-app prompt rules" is now "Per-app rules".
Fixed
Long recordings failed with local models, and local recognition stopped working afterwards.
A long recording was passed to the model in one piece. On integrated graphics this could exceed
the time Windows allows the GPU, the GPU was reset, and only restarting SayIt helped. Recordings
are now split at pauses into pieces of up to 30 seconds, which also makes long recordings faster.
If the GPU still fails, SayIt switches to the CPU until it is restarted. In local mode, the wait
for a live recording's result now grows with its length, up to 3 minutes (it used to be at most
45 seconds).
Inserted text showed up in the history of clipboard managers such as Ditto. SayIt only set
the marker that Windows clipboard history checks. It now also sets the one third-party clipboard
managers commonly use.
Re-transcribing on the History page failed without any message. The reason now appears under
the record, including when the audio file has already been cleaned up.
Date groups in History and Favorites were out of order. After a new month began, the newer
month could appear further down, and the English UI sorted months alphabetically.
Long Qwen Omni recordings (about a minute and a half or more) failed to transcribe. The whole
recording was sent at once, which on a fast connection exceeded the realtime endpoint's upload
limit and the server closed the connection. Recordings under a minute are still sent at once;
longer ones send their second half more slowly (about half a second extra for 99 seconds, about
2 seconds for 3 minutes), and if the limit is still hit, SayIt retries once at a slower pace.
Idle CPU usage crept up after a few dictations. Audio processing for the ready tone kept
running in the background until you quit. It now pauses after each tone.
Images and files on the clipboard were lost after inserting text. Only text used to be
restored. The whole clipboard is now put back (with "Restore the clipboard after inserting" on).
Hotkeys occasionally stopped working until SayIt was restarted. Windows can quietly remove
the keyboard hook when the system is busy; SayIt now notices and reinstalls it. An error while a
recording was starting also no longer leaves every later hotkey dead.
Voice input stopped working when the main window's page crashed or failed to load. The page
now reloads, or the window is recreated, automatically; if that also fails, a message asks you to
restart SayIt.
SayIt would not start if its local database was damaged. The damaged file is now renamed and
kept, SayIt starts with a fresh database, and a message explains what happened.
Total dictation time under Usage did not match the home page.
Improved
Lower background usage. The mouse is only watched when a hotkey uses a mouse button, window
switches no longer trigger extra work, server mode sends audio in batches, and the overlay
refreshes about half as often while recording.
Less memory used while the main window is in the tray.
Long recordings and diagnostics packages are handled faster: audio and screenshots are passed
between the interface and the background service directly instead of being re-encoded.
Interrupted update downloads resume the next time SayIt checks, and updates are no longer
installed while Windows is shutting down.
Settings that depend on AI cleanup say when they have no effect. With AI cleanup off, context
awareness and "Send hotwords to the AI prompt" are greyed out, and the services on the AI
providers page are marked "Not in use". The same mark appears in server mode while the server is
doing the cleanup.
WebDAV backups that cannot reach the server now say so plainly, with the original error under
"Details".
Security
HTTPS certificates are checked again. Self-signed certificates are accepted only for a server
address you entered yourself.
Update installers are downloaded only from the same origin as the update server, and their
location and hash are checked before installing.
Importing a backup no longer overwrites settings that belong to this computer, and audio in a
backup must be .wav files.
File operations are limited to their own folders, and a Content Security Policy is enabled.
