🚀 Beta Test – Murmure 1.11

[Kieirra]

Hello everyone,

The beta for Murmure 1.11.2 is open. It is an open beta, anyone who wants to test the pre-release is welcome, no sign up needed.

Last time the test plan was long and hidden behind a link, and nobody went through it. So this time everything is right here, and the essentials take about five minutes.

✅ The essentials, about 5 minutes

• Dictate a sentence like you normally do, and check the text lands correctly
• Select text in any app, press Ctrl+Alt+Shift+1, and check the prompt runs on your selection (this is the new Transform gesture, set up mode 1 in LLM Connect first if you never did)
• Add a two-word entry to your dictionary, like a first and last name, then dictate it
• Turn on the volume reduction in Settings > System, play some music, then dictate

Even one checked box helps.

➕ If you have more time

• Change one of your shortcuts in Settings, then use it
• Copy an image, then dictate, and check the image is still in your clipboard
• Dictate something long, over a minute, and check nothing is missing at the end
• Open the dictionary and check the word count indicator (green up to 50 words, yellow from 51 to 100, red above 100)

💻 Only the line matching your setup

• macOS: bind the fn key (the one with the globe icon) to a shortcut, then use it
• Linux Wayland: dictate a sentence with accented characters in direct insert mode
• Arch or CachyOS: install the .pkg.tar.zst package and start the app

📦 Downloads

All binaries are on the pre-release page: https://github.com/Kieirra/murmure/releases/tag/1.11.2

• 🐧 Linux
  • Murmure_amd64.deb (Debian, Ubuntu)
  • Murmure_amd64.rpm (Fedora)
  • Murmure_amd64.AppImage (AppImage)
  • Murmure_amd64.pkg.tar.zst (new, Arch and CachyOS)
• 🍎 macOS
  • Murmure_aarch64_darwin.dmg (Apple Silicon)
  • Murmure_x86_64_darwin.dmg (Intel)
• 🪟 Windows
  • Murmure_x64.msi (.msi)
  • Murmure_x64-setup.exe (.exe)

✨ What's new in 1.11

LLM Connect

• Transform: select text in any application, press Ctrl+Alt+Shift+1 to Ctrl+Alt+Shift+4, and the mode's saved prompt runs on your selection, with no dictation at all
• Remote providers: you can now type a model name by hand, and a failed connection test no longer locks the provider, so servers without a /models endpoint (such as the Claude API) work
• The temperature parameter is no longer sent to remote servers, which fixes the 400 Bad Request returned by OpenAI GPT-5 models

Dictionary

• Light redesign, with a word count indicator: green up to 50 words, yellow from 51 to 100, red above 100
• Two-word entries are now accepted, so expressions with a space or a hyphen can be added
• All characters from Parakeet's vocabulary are allowed, not only letters
• Export your dictionary as a .txt file, and import one with a documented format

Shortcuts

• Pause and ScrollLock can be used as shortcut keys on Windows and Linux
• The fn key, marked with a globe icon on recent Macs, can now be bound on macOS
• Keys are read from the native backend, so F13 is no longer shown as "Unidentified" and letter labels match your real keyboard layout on X11

Audio

• The system volume can go down while you record, so you hear yourself better
• Microphone level detection adapts to your gain and to background noise
• Each chunk is padded with silence before transcription, which fixes results that were silently cut

Linux

• New pacman package for Arch based distributions, including CachyOS
• Wayland: accented characters are typed natively in direct insert mode
• Murmure now appears under Utility in application menus

API and CLI

• The local API accepts audio of any length, as long as the request stays under 100 MB, and stops transcribing as soon as the client disconnects
• New --hidden flag to start Murmure without showing the window

Other

• Logs no longer grow forever: the file is reset when it goes above 1 MB while the app runs, and a new Off level disables logging completely
• A copied image stays in your clipboard when you dictate, instead of being replaced
• The wake word listener no longer restarts in a loop when the microphone is unavailable

🐛 Found something broken?

Just reply here with what broke and on which OS. That is already enough. Steps to reproduce and the log file are welcome if you have them (enable debug mode in Settings > System, then reproduce the bug).

Full plan in the docs: 🇬🇧 https://docs.murmure.app/beta-testing/ · 🇫🇷 https://docs.murmure.app/fr/beta-testing/

❤️ Thank you

Thanks to everyone who takes a few minutes for this, it is what makes the stable release solid.

---

Edit — the binaries were rebuilt as 1.11.1. Fixed since the first beta:

• The start and stop sounds were far too quiet, they are much louder now, and a slider in Settings > System sets their volume ([Feature request] Make the notification sound louder #426)
• Recording no longer makes the KDE volume popup appear
• Bluetooth microphones are listed again on Linux (fix(audio): list bluetooth microphones on linux #424)
• The microphone icon no longer blinks every ten seconds while the wake word listener waits in silence (fix(wake-word): résolution du clignotement de l'icône micro sous Linux #423)

---

Edit 2 — the binaries were rebuilt as 1.11.2. Fixed since the previous beta:

• Transform now waits until Ctrl, Alt and Shift are physically released before copying your selection, so it no longer fails with "no text selected" on Windows, X11 and macOS (thanks @laurentkempe)
• The audio output stays open while you record, so the stop sound is no longer played on a device that already went back to sleep on dictations longer than a minute (thanks @tomsquest)
• On X11, the keycodes borrowed to type characters missing from your layout are released when the app quits, so the keymap no longer fills up run after run until both Ctrl+V and Direct insert stop working

[tomsquest]

edit: rebooted my system (previous paste issues maybe related to a system update)

Tested on Kubuntu 26.04.

The essentials

• [OK] Dictate a sentence like you normally do, and check the text lands correctly
• [NOT TESTED] Select text in any app, press Ctrl+Alt+Shift+1, and check the prompt runs on your selection (set up mode 1 in LLM Connect first if you never did)
• [OK] Add a two-word entry to your dictionary, like a first and last name, then dictate it
• [OK] Turn on the volume reduction in Settings > System, play some music, then dictate

(this was not working, but it was a paste issue on my system, working fine now).

KDE displays a popup when sound volume change or is muted, which is kindof disturbing (2 "popups": sound volume and Murmure widget).

No problem of focus stealing.

If you have more time

• [NOT TESTED: wayland, using murmure --transcription-bound shortcut] Change one of your shortcuts in Settings, then use it
• [NOT TESTED: I use Klippy, aka multiclipboard, so images are kept] Copy an image, then dictate, and check the image is still in your clipboard
• [OK] Dictate something long, over a minute, and check nothing is missing at the end
• [OK] Open the dictionary and check the word count indicator (green up to 50 words, yellow from 51 to 100, red above 100)

Only the line matching your setup

• [n/a] macOS: bind the fn key (the one with the globe icon) to a shortcut, then use it
• [OK] Linux Wayland: dictate a sentence with accented characters in direct insert mode
• [n/a] Arch or CachyOS: install the .pkg.tar.zst package and start the app

[tomsquest]

@Kieirra Tested Handy, just to make sure, and Handy does not display any Kde popup when muting sound (there is no option to reduce the volume). This has 2 advantages: no KDE popup "sound muted" and the notification sounds are still played.

[Kieirra]

Yes, thanks for testing. I have an idea for removing the KDE popup, so that seems to be the easy part.

So, is this the only problem in the end?

[tomsquest]

Yes, i think everything else worked fine (I did not test the advanced features like llm connect)

…

On 19 August 2026 17:50:16 CEST, "🇫🇷 Al1x-ai" ***@***.***> wrote: Yes, thanks for testing. I have an idea for removing the KDE popup, so that seems to be the easy part. So, is this the only problem in the end? -- Reply to this email directly or view it on GitHub: #419 (reply in thread) You are receiving this because you commented. Message ID: ***@***.***>

[Kieirra]

OK perfect

[tomsquest]

Hi @Kieirra,

Thanks for the update!

Testing 1.11.1 on my laptop this time (still Kubuntu 26.04, but another device).

My report:

1. Start or End sounds are not always emitted (heard).

Sometimes it's the start sound that is skipped, sometimes that's the end sound.
Tried at 80% (default) and 100%.

Here is the log, in debug, but I did not see anything related to sound.
murmure.log

2. No animation of the widget

The widget animation is stuck while recording.
Sometimes it moves when starting a recording (you can see in the screencast below), but no animation at all while recording.
The final animation (when I speak for a longer time) is displayed however.

I am pretty sure it's not new and it was already the case before (I just forgot about it, I mainly use Murmure on my main desktop computer).

Here I recorded twice, the widget does not animate:

[Kieirra]

Thank you again for the two reports.

About the sounds: I had the same problem on my Linux desktop. It does not come from Murmure: Linux suspends the sound card after a few seconds of silence to save power, and the wake up delay eats the beginning of short sounds. I cannot keep the output open all the time as a workaround, it drains the battery of Bluetooth devices (#368) but you can configure it with something like that :
echo "options snd_hda_intel power_save=0" | sudo tee /etc/modprobe.d/audio_powersave.conf
(Just be aware that this may affect Bluetooth devices or increase battery usage.)

One part is on my side though: on dictations longer than a minute, Murmure closes its output during the recording, so the end sound plays on a cold device. That one I can fix.

About the widget animation: probably the same kind of thing, it depends on the machine more than on Murmure. It looks like a known rendering issue of WebKitGTK (the system webview Murmure relies on under Linux) on some GPUs, especially laptops on Wayland. That would explain why your desktop is fine and your laptop is not.

[laurentkempe]

Hi @Kieirra Thx for the new version, giving back a bit testing it! All answers written by Murmure 🎉

✅ The essentials, about 5 minutes

• Dictate a sentence like you normally do, and check the text lands correctly
• Select text in any app, press Ctrl+Alt+Shift+1, and check the prompt runs on your selection (this is the new Transform gesture, set up mode 1 in LLM Connect first if you never did)
  It was never working. I tried with push to talk and toggle to talk
  https://github.com/user-attachments/assets/e374805b-d8f1-4bed-aaf0-510734580590
• Add a two-word entry to your dictionary, like a first and last name, then dictate it
  Worked for my first name but not for my last name.
• Turn on the volume reduction in Settings > System, play some music, then dictate
  It works really nicely. When the sound is very loud and you stop dictating, then it turns comes back very loud directly. I would have something that increases the volume in some steps.

➕ If you have more time

• Change one of your shortcuts in Settings, then use it
• Copy an image, then dictate, and check the image is still in your clipboard
• Dictate something long, over a minute, and check nothing is missing at the end
• Open the dictionary and check the word count indicator (green up to 50 words, yellow from 51 to 100, red above 100)
  15 words in green. I didn't test with more.

[Kieirra]

Thank you a lot for the tests

Worked for my first name but not for my last name.

I think it works, but it depends on the pronunciation. The algorithm replaces it correctly when it understands the phonemes but isn’t sure about the words.

For example, if I pronounce "Laurent Kempe" like "Laurent Kempé", it works on my side. But yes, names are difficult because they often have their own phonetic rules, which aren’t tied to the language being used, I guess.

It was never working. I tried with push to talk and toggle to talk

This is concerning. Okay, I’ll investigate. Thank you for the logs.

[Kieirra]

It was never working. I tried with push to talk and toggle to talk

I think I found the cause. After the shortcut, Murmure sends Ctrl+C to copy your selection. If Ctrl, Alt or Shift are still held down at that moment, the copy fails and you get the "select text" message.

Can you try with a single key like F8 as the Transform shortcut? It triggers on release, so give it a short press and release it right away. If that works, it confirms the cause. I will make the next version wait until all keys are released before copying.

[laurentkempe]

For example, if I pronounce "Laurent Kempe" like "Laurent Kempé", it works on my side. But yes, names are difficult because they often have their own phonetic rules, which aren’t tied to the language being used, I guess.

I will try again with "Laurent Kempe", I had mixed results

then a bit better when I got used to it

For sure it is not the way I am used to say my name, but as you said this is really tricky.

Can you try with a single key like F8 as the Transform shortcut? It triggers on release, so give it a short press and release it right away. If that works, it confirms the cause. I will make the next version wait until all keys are released before copying.

It works like that (F8), and it make sense that keeping the key pressed you get the experience I had.
I have to be honest, I think I never understood the way it works. I select some text, press shortcut, and was saying translate to French. And in fact, you apply the prompt, so I had to add the Translation prompt but then it is fixed to English. I was expecting that what I am saying is what will be done on the selected text.

[Kieirra]

lol, I probably mispronounce your name a lot, but it seems to work.

One thing though: test the dictionary with a real sentence, not just "Laurent Kempe" on its own.

A dictionary word only gets boosted if the model already ranks it close to what it heard, and the decoder is recurrent, so whatever comes before your name changes that ranking. A name with nothing in front of it is the hardest case you can give it.

And even then, Murmure only swaps a word for a close dictionary entry when the model was unsure about it in the first place.

---

I have to be honest, I think I never understood the way it works. I select some text, press shortcut, and was saying translate to French. And in fact, you apply the prompt, so I had to add the Translation prompt but then it is fixed to English. I was expecting that what I am saying is what will be done on the selected text.

Yes, a lot of people miss this, that's why I added this:

Hover the "?" and you get a tooltip for each gesture.

Mode	Tooltip
Command
Transform
Dictate (standard LLM Connect)

Short version:

• 1 is your voice as the instruction,
• 2 is the prompt on a selection,
• 3 is the prompt on your dictation.