French not working well (large-v3 installed, Language Auto)

[jfbourdeau]

I installed large-v3 ( I tried large-v3-turbo too)

language is set to Auto

Translation is disabled

English is working fine but not French

Is there anything I could have missed ?

Do we set it to auto or "auto" ?

[whisper]
model = "large-v3-turbo"
language = "auto"
translate = false

[peteonrails]

Yes, that configuration should net you some pretty good results when speaking French. If it doesn't, you can feel free to open up an issue and we can start tracking it and working on it. My French is pretty terrible, so I might need your help debugging any optimizations that we make.

[jfbourdeau]

when creating an "issue", it's asking to run voxtype -vv

My question ( noobie ) :
If VoxType is already running on Omarchy, can I voxtype -vv to provide the log ?

Or I must kill/Stop the processes and then voxtype -vv and dictate my french words

[peteonrails]

Good question! The -vv flag only affects a new daemon instance. It won't add logging to the one that's already running.

For debugging, the easiest approach is to temporarily stop the service and run voxtype manually in a terminal:

# Stop the running service
systemctl --user stop voxtype

# Run with verbose logging (you'll see output in this terminal)
voxtype -vv daemon

Now test your French dictation. You'll see real-time logs showing:

• Language detection results (auto-detected language: fr)
• The raw transcription before any processing
• Any errors or warnings

When you're done debugging, press Ctrl+C to stop, then restart the service:

systemctl --user start voxtype

If you share the relevant log output, we can figure out what's happening with French.

[jfbourdeau]

Zzzz I uninstalled it, reinstalled it, checked my config and " it is now working in french and english !!! " No idea why

But I am having a random problem, Only a small sentence (few words) make my computer work hard, the fan spin during 20 seconds and I get the dictation results and the FAN speed is going down..

My PC
Is this some GPU Voxtype tweaking I need to do ?

[peteonrails]

No, that is a problem with the selection of context window sized for Whisper. This has been fixed in a recent pull request and will be in the next release. Instead of using the default 30 second context window, we will make Whisper use a context window that is proportional to the size of the recording.

I'm going to tag a build after 5 o'clock p.m. Eastern Time today, and then the binaries should be updated around 6 p.m. With the way that the AUR updates, it should be available in Omarchy by about 6.30 p.m. EST (GMT-5)

[jfbourdeau]

WOw cool !! tks
👍👍

[peteonrails]

@jfbourdeau The latest release is out with optimized Whisper context windows.

[jfbourdeau]

I just unsinstalled and reinstalled VoxType as I had a ydotool error. Is it important ? See below :

utput Chain:
Display server: Wayland (WAYLAND_DISPLAY=wayland-1)
wtype: ✓ installed (/usr/bin/wtype)
ydotool: ⚠ installed (/usr/bin/ydotool), daemon not running
wl-copy: ✓ installed (/usr/bin/wl-copy)

[peteonrails]

That is not a problem. Voxtype will use wtype by default. That is simply a warning because it is one of the fallback methods. I'll look at making that message a little bit more clear.

[jfbourdeau]

French is working.
When using large-v3-turbo or large-v3 it takes over 10 seconds to convert what I say to words
With something like " This is a cat ".
If I can help let me knnow what to test.
When I just installed it and use the default model and speak english, then it's very fast !

But as soon as I use one of the 2 model that I think I need to use for MultiLingual, Computer works hard, fan is going up and I get the result after 10 secs

2026-01-16T14:20:41.311285Z TRACE State file updated: transcribing
2026-01-16T14:20:41.311373Z DEBUG Transcribing 1.92s of audio (30725 samples)
whisper_backend_init_gpu: no GPU found
whisper_init_state: kv self size = 83.89 MB
whisper_init_state: kv cross size = 251.66 MB
whisper_init_state: kv pad size = 7.86 MB
whisper_init_state: compute buffer (conv) = 36.15 MB
whisper_init_state: compute buffer (encode) = 212.31 MB
whisper_init_state: compute buffer (cross) = 9.27 MB
whisper_init_state: compute buffer (decode) = 99.12 MB
2026-01-16T14:20:41.346868Z INFO Audio context optimization: using audio_ctx=160 for 1.92s clip (formula: 1.92s * 50 + 64)
whisper_full_with_state: auto-detected language: fr (p = 0.998342)
2026-01-16T14:20:56.672276Z INFO Transcription completed in 15.36s: "Un petit chat."
2026-01-16T14:20:56.674814Z INFO Transcribed: "Un petit chat."
2026-01-16T14:20:56.675479Z DEBUG ydotool does not support --key-hold flag, using --key-delay only
2026-01-16T14:20:56.676044Z DEBUG Running: wtype -d 1 -- "Un petit chat."
2026-01-16T14:20:56.755976Z DEBUG Text output via wtype
2026-01-16T14:20:56.756073Z TRACE State file updated: idle

[peteonrails]

whisper_backend_init_gpu: no GPU found

This is the culprit. Voxtype is not finding your GPU.

What kind of GPU do you have in your machine? If you run voxtype setup gpu and restart the daemon, does it recognize your GPU on startup?

[jfbourdeau]

I just changed / switch to base.en and I still get over 10 secs (without uninstalling)

So the problem is weird... Brand new installation, it's fast.

I change the model, after it's over 10 sec whatver model I use

and french works

[jfbourdeau]

In case it helps
Over 10 s ecs for 3 or 4 words (French or English, the same)
JF
And as mentionned, If I uninstall, reinstall, don't change anything, it's very very fast. (English test)

~ ✗ voxtype config
Current Configuration

=====================

[hotkey]
key = "SCROLLLOCK"
modifiers = []
mode = PushToTalk

[audio]
device = "default"
sample_rate = 16000
max_duration_secs = 60

[audio.feedback]
enabled = false
theme = "default"
volume = 0.7

[whisper]
model = "base.en"
language = "auto"
translate = false

[output]
mode = Type
fallback_to_clipboard = true
type_delay_ms = 1
pre_type_delay_ms = 0

[output.notification]
on_recording_start = false
on_recording_stop = false
on_transcription = false

[status]
icon_theme = "emoji"
(resolved icons: idle="🎙\u{fe0f}" recording="🎤" transcribing="⏳" stopped="")

[integration]
state_file = "auto"
(resolves to: "/run/user/1000/voxtype/state")

Output Chain:
Display server: Wayland (WAYLAND_DISPLAY=wayland-1)
wtype: ✓ installed (/usr/bin/wtype)
ydotool: ⚠ installed (/usr/bin/ydotool), daemon not running
wl-copy: ✓ installed (/usr/bin/wl-copy)

→ Text will be typed via wtype (CJK supported)

[jfbourdeau]

✗ voxtype setup gpu
=== Voxtype Backend Status ===

Active backend: CPU (AVX2)
Binary: /usr/lib/voxtype/voxtype-avx2

Installation mode: tiered (pre-built)

Available backends:
CPU (AVX2) - active
CPU (AVX-512) - installed
GPU (Vulkan) - installed

GPU detected: Advanced Micro Devices, Inc. [AMD] Device 43d1 (rev 01)
Vulkan runtime: installed

To enable GPU acceleration:
sudo voxtype setup gpu --enable

[jfbourdeau]

Radeon RX 470/480/570/570X/580/580X/590] (rev e7)

[jfbourdeau]

enable GPU seems to have fixed everything
~ ❯ sudo voxtype setup gpu --enable
[sudo] password for yoda:
Switched to GPU (Vulkan) backend.

Restart voxtype to use GPU acceleration:
systemctl --user restart voxtype

~ ❯ systemctl --user restart voxtype

~ ❯ This is a test.La belle petite fille.Le beau petit chat.I love this cat.This is a nice house.

[jfbourdeau]

May be one last unrelated question (Tks again btw )
I enabled punctuation and I get this : The little girl, comma, meets a cat, comma, and tell him hello.

[peteonrails]

That's unusual. What model are you using?

[jfbourdeau]

[ 5] large-v3
[ 6] large-v3-turbo

Both are doing the same.

This is an example, comma, of using punctuation, comma.

[peteonrails]

Can you try disabling spoken_punctuation in your voxtype/config.toml?