Is it possible to get a voice activation/stop option in "murmure" ?

[OlafausBerlin]

First of all, thank you for developing this software. It makes life so much easier.

There is the software "Talon Voice" that can also transcribe speech to text and you can control your computer with commands. It does not use the "parakeet" model that i prefer for German language transcription and it is complicated to use in general.

When you start "Talon Voice" it is listening in the background, but does not record anything until you say "wake up".
Then it starts to record your speech.
When you say "go to sleep", "Talon Voice" stops recording and goes back to listening mode.
So there's no need to press a key.

In "murmure" i've tried to implement a custom rule to end recording like "strg + space"(German keys) when saying "ende" in murmure, but so far it does not allow "strg" as a key.
Maybe it is already possible to use that key combination in custom rules, but i don´t know how.
"Enter" is allowed and space can be uses as well in custom rules.

Is it possible to get a voice activation/stop option in "murmure" ?

[Kieirra]

Thank you for the kind words and for the detailed feedback.

I didn't know Talon Voice before, I'll take a look at how it works.

For voice activation, you're not the first one to request this so I'm going to work quicly on a proof of concept to see how feasible it is, the goal being to trigger recording with a wake word while keeping CPU usage as low as possible instead of using shortcuts. I'll keep you posted on how it goes.

[thiswillbeyourgithub]

Just a thought: i've created an app that used STOP to split o tinuous audio on the fly in the past and can tell you that it definitely needs code that removes silences, at the very least strips ends. So: related to #166

[Kieirra]

Just a thought: i've created an app that used STOP to split o tinuous audio on the fly in the past and can tell you that it definitely needs code that removes silences, at the very least strips ends. So: related to #166

Thanks for sharing your experience with Voice2Anki.

The approach I have in mind is different though, I'd use a separate, lightweight wake word detector (not Parakeet) that just listens for a keyword to start/stop the recording like this https://crates.io/crates/pv_porcupine).

It simply replaces the keyboard shortcut as a trigger. The rest of the pipeline stays the same. So silences aren't more of a concern than they are today.

The only thing to handle is the stop word captured in the audio. Either trim the last few milliseconds (matching the wake word duration) before sending to Parakeet, or simply strip it from the transcription output. Both are straightforward but in need to POC it too see if it's really that simple or if I'm missing something.

[Kieirra]

Just seeing that https://crates.io/crates/pv_porcupine requires an online connection, but fully local alternatives exist like https://crates.io/crates/oww-rs. I need to dig a bit more, it seems to use ONNX too, so I'm not 100% sure about using it instead of Parakeet. Well... I'll check all of that when I have some time.

[Kieirra]

I haven’t finished the implementation yet, but the results are already promising.

#178

The feature can be enabled in Settings by entering a simple, easily recognizable keyword (e.g. “beta”).

Here’s the current behavior:

1. The app continuously listens to the microphone volume in the background.
2. When it detects speech for about 200 ms, it records a short audio segment (up to 5 seconds), keeping a small buffer so the beginning of the sentence is not lost.
3. This short segment is transcribed locally with Parakeet to check if the keyword was spoken.
4. If the keyword is detected, recording starts automatically.
5. Recording stops after 1.5 seconds of silence.

This is still a pure POC, there’s no dedicated UX/UI yet. I haven’t decided whether to:

• Create a dedicated voice assistant page, or
• Keep the current approach and simply allow users to configure shortcuts and enable a wake word when activated.

Next steps:

• Continue evaluating performance impact
• Ensure it does not introduce regressions or stability issues
• Define a clear and simple UX/UI to enable and configure it
• Refactor the code

More updates soon.

[thiswillbeyourgithub]

Nice.

1. Is it already possible to cancel a recording or not? Because I can imagine annoyance like you click on some video that plays sound that triggers the wake word. So it would be nice to still have a shortcut to cancel like shift+esc or clicking on the overlay.
2. Adding a shortcut to enable / disable listening for the wake word would be nice if you haven't done this already.
3. Hotword is usually a single specific word because it relies on a model trained specifically on that word. But if you rely on an ASR model it will use context to parse words so "beta" might become "d'état", "jarvis" could become "j'ai une vis", "okay, alex à" etc. Hence, I have a hunch that it would be best to allow adding several trigger words, or even regexes.

[Kieirra]

Is it already possible to cancel a recording or not? Because I can imagine annoyance like you click on some video that plays sound that triggers the wake word. So it would be nice to still have a shortcut to cancel like shift+esc or clicking on the overlay.

Yes, in 1.8.0, you can cancel a recording with "Escape" by default.

Adding a shortcut to enable / disable listening for the wake word would be nice if you haven't done this already.

Currently, you can enable or disable it in the system settings, but there is no keyboard shortcut. I already have quite a few shortcuts in the app. I’d prefer not to add any more.

Hotword is usually a single specific word because it relies on a model trained specifically on that word. But if you rely on an ASR model it will use context to parse words so "beta" might become "d'état", "jarvis" could become "j'ai une vis", "okay, alex à" etc. Hence, I have a hunch that it would be best to allow adding several trigger words, or even regexes.

Yes, it works well with words like "beta", but you’re right, "Alix" can become "Alice", that kind of thing.

I think using the same dictionary algorithm or a more eager “near distance match” algorithm could solve this, so it stays transparent for users.

And yes, you can already add more than one word, like "ok google".

("jarvis" could become "j'ai une vis" :D)

[Kieirra]

Done, will be available in 1.8.0.

Works like a charm, if you want to see screenshots : #178