Each code works once if you lose your authenticator. Store them somewhere safe; they are shown only now.
Voice cloning · Qwen3-TTS
Read one short text. Get your own voice.
Pick a language, read the teleprompter text out loud for about 10 seconds, and voicekit
builds a voice profile from it. Your recording never leaves this computer.
Set up
Name your voice, choose the language you will speak, and turn on your microphone.
Lowercase letters, digits, “-” and “_”.
LanguageMicrophone
Quiet room, about 20 cm from the microphone, no headphones leaking sound.
Read aloud
Press record, wait for the countdown, then read the text calmly and naturally — exactly as written.
Text 1 / 4REC
0.0 sIdeal length: 6–15 seconds
Check & train
Listen back. If you misread a word, record again — the text must match the audio exactly.
Training
Status updates appear here live.
✓ Voice “” is ready
Try it in the terminal:
Training failed
Show details
Your voices
Admin
API keys
Keys let your apps use this server. Give each app its own key with only the scopes it needs.
No login is configured yet, so this panel is open to this computer only.
Creating the first key turns on login. Create an admin account first on the server:
pdm run voicekit users add <name> --admin
+
Create a key
The key is shown once. Copy it into your app's server-side configuration.
Scopes
✓ Key “” created
Copy it now. Only a hash is stored, so it cannot be shown again. Send it as
Authorization: Bearer <key> from your app's backend, never from browser code.
Admin users
New accounts get your permissions. Passwords must be at least 10 characters.
Invite a user by email
The recipient chooses their own password using a link valid for 48 hours.
Set an account email
Existing accounts need an email address before they can request a password reset.
Send a password reset link
Send a one-hour reset link to the account’s saved email address.
Multi-factor authentication
Require an authenticator app for all user accounts. API keys are not affected.
Changing this setting signs out account sessions. Users without an authenticator will set one up at their next sign-in. Keep recovery codes safe.
Active keys
Revoking takes effect immediately for API calls and signed-in sessions.
Admin
Ollama models
Download, back up and choose the model used by the AI writer.
GPU fit is an estimate from model download size and a single GPU’s total memory. Runtime, context, other workloads and Ollama’s GPU configuration affect actual fit. Large models may run partly on CPU.
Preparing
DownloadBackup & verifyReady
✦
Test the selected model
Chat with the same model used by the API. Try the sample prompt on different models to compare speed.
Selected model:—
✧Start a conversationAsk a question or run the sample to compare models.
Response reached the token limit and may be incomplete. Increase the limit or ask the model to continue.
Generating a reply…
↗
Model comparisons
Last test on this browser. Speeds vary with prompt, context, GPU load and whether the model was already loaded.
◇
Installed models
Select a model for the API and AI writer before loading it, or make an offline backup.
↓
Suggested downloads
Downloads are automatically backed up and verified before completion.
Offline backups
Backups contain verified model files and their exact manifest. To restore after model loss, stop Ollama, copy the backup blobs and version’s manifest into Ollama’s model directory, fix ownership, and restart Ollama. Keep another copy on a separate drive for protection against disk failure.
Admin
Email settings
Configure delivery for invitations, password resets and health alerts.
Credentials are stored unencrypted in an owner-only file on this server. SMTP and IMAP connections use TLS. Configure an HTTPS public URL for account links.
Mail servers
Leave an existing password empty to keep it. IMAP is optional and used to verify mailbox access; outgoing mail uses SMTP.
Admin
GPU energy
Track measured GPU power, energy and estimated electricity cost over time.
GPU power is sampled every 5 seconds. “During backend work” estimates energy while Voicekit jobs or text requests run; it includes idle draw and other workloads on the GPUs. CPU and the rest of the server are not measured.
Totals include only measurements while Voicekit is running; missing periods are excluded.
Live GPU power
—
Waiting for readings…
Energy over time
Daily totals in your local time. Only days with measurements are shown.
All GPU energyDuring backend work
No energy readings yet.
Show daily values
Date
All GPU energy
During backend work
Average power
Estimated cost
Energy by GPU
Measurements from the last 30 days.
Electricity price
A single flat rate is applied to all displayed history. Add taxes and fees to the rate if needed.
Sign in with an administrator account to change the price.
Admin
System health
Check the server, services, GPU and CUDA. Refresh to take a new snapshot.
Services
GPU & CUDA
CUDA availability is checked using this app’s PyTorch installation.
Admin
Activity log
Sign-ins, API calls, key and voice changes, job results. Texts, passwords and keys are never recorded.
Time
Event
Who
IP
What
Details
No events match.
Text to speech · your voices
Type it. Hear it in your voice.
Write with AI (optional)
Describe the text you need. The draft goes into the editor in the next step, where you can change anything. Already have a text? Just continue.
Try for example:
Enter to send · Shift+Enter for a new line · AI drafts can contain mistakes — check names, numbers and times.
Text & voice
Adjust the text until it is exactly what should be said, choose the voice, then create the audio.