Releases: soupslurpr/Transcribro
0.3.0
Transcribro's keyboard now adapts the transcript to the surrounding and selected text, allowing for easier editing without having to go back to other keyboards as much. Check out the description of the PR at #39 for an example of using it and information on how to make use of it properly. Even without reading, its been made seamless to use.
The keyboard also plays audio feedback when recognition gets toggled. Keep in mind that it will only play if your ringer is set to sound rather than vibrate or mute.
Next, the Whisper model being used was switched from base-en-q5_1 to tiny-en-q8_0, which means computing the transcripts is now a lot faster! There doesn't seem to be a noticeable quality loss in my limited testing.
The last major addition is an ACTION_RECOGNIZE_SPEECH activity, which allows other apps to ask Transcribro to launch on top of them, transcribe audio from the microphone, and then return the transcript to them.
Chromium, for example, uses this (by pressing the microphone button at the right of its search bar as of writing) to search, and seems to tell the search engine that speech was used to search. For Google search, a voice tells you a summary or such of the information you requested, acting sort-of like Google assistant, which is pretty cool.
Finally, the privacy policy was updated.
If there are any issues, please don't hesitate to create an issue in GitHub (first make sure to search in case a similar issue already exists), and/or join the Matrix rooms (check https://github.com/soupslurpr/transcribro?tab=readme-ov-file#community).
Enjoy!
Notable Changes
- Audio feedback when recognition gets toggled by @soupslurpr in #24
- update dependencies by @soupslurpr in #26
- switch to tiny.en-q8_0 whisper model by @soupslurpr in #29
- Switch to pinned top bar with no custom windowInsets by @soupslurpr in #30
- Automatic navigation animations by @soupslurpr in #31
- Add ACTION_RECOGNIZE_SPEECH activity by @soupslurpr in #32
- Update whisper.cpp to af5833e29819810f2d83228228a9a3077e5ccd93 by @soupslurpr in #34
- Ensure audio inputs are a certain length to prevent them going unprocessed by @soupslurpr in #35
- Update privacy policy by @soupslurpr in #36
- Keyboard: When the transcript is an empty string, don't commit it by @soupslurpr in #37
- Dependency bumps by @soupslurpr in #38
- Keyboard: Adapt output transcript capitalization and punctuation to the surrounding and selected text by @soupslurpr in #39
Full Changelog: v0.2.1...v0.3.0
0.2.1
The recognition service was somewhat reworked to be more robust and fix some bugs (transcribing longer works again now) and it actually results in increased speed and accuracy because we can now remove some unnecessary padding to the end of the speech which might've been causing hallucinations such as "Thank you." and "Thanks for watching". Those hallucinations could still happen in certain circumstances, but testing has shown a heavy reduction of them!
Notable Changes
- Make the recognition service more robust by @soupslurpr in #23
Full Changelog: v0.2.0...v0.2.1
0.2.0
The latest Transcribro update enhances transcription accuracy with an upgraded speech recognition model and a manual check to remove a specific frequent error. Audio data gets deleted after transcribing when using Auto-Stop Recognition instead of doing it after recognizing, resulting in more memory efficiency. New action and return keys on the Voice Input Keyboard are here, along with a slight change in the privacy policy for the new Auto-Send Transcription feature, which automatically presses send after transcribing and can be enabled in settings. Bug fixes include resolving Voice Input Keyboard malfunctions and removing unwanted space before transcriptions.
For more details, please look below.
Notable Changes
- Increase speech pad ms by a second to workaround VAD sometimes stopping too early by @soupslurpr in #5
- use base.en-q5_1 model instead of tiny.en-q8_0 to increase transcription quality and heavily reduce hallucinations by @soupslurpr in #6
- remove " ." to reduce hallucination that sometimes appears at the end of the transcript by @soupslurpr in #7
- more efficiency and race condition proofing by @soupslurpr in #10
- action and return key and don't add space character if there are no characters before the cursor or if the first character before the cursor is a newline by @soupslurpr in #12
- update privacy policy to account for planned automation features by @soupslurpr in #14
- fix Voice Input Keyboard sometimes silently breaking by @soupslurpr in #15
- Auto-Send Transcription (#2) by @soupslurpr in #16
- update dependencies by @soupslurpr in #17
Full Changelog: v0.1.1...v0.2.0
0.1.1
Quick bugfix to cancel recognition breaking the recognition service😅
Notable Changes
- fix recognition not working after cancelling by @soupslurpr in b5f6f80
Full Changelog: v0.1.0...v0.1.1