Frequently Asked Questions

General

Is Conjure free?

Yes. Conjure is fully open-source under the AGPLv3 license. The desktop app, local whisper transcription, and all features are free to use. There is no subscription required.

If you use cloud providers (OpenAI, Groq, etc.), you pay those providers directly for API usage. Many providers offer generous free tiers -- Groq's free tier is sufficient for most users.

Is Conjure safe to use?

Conjure is open source -- you can audit every line of code. The app runs locally on your machine. Audio and text are only sent to providers you explicitly configure. There is no telemetry, analytics, or data collection by Conjure itself.

Does Conjure work offline?

Yes. With local whisper selected as your transcription provider, Conjure works entirely offline. No internet connection is needed for dictation, voice commands, dictionary, or any local features.

Cloud providers (for transcription, post-processing, TTS, or agent mode) require internet access.

What languages are supported?

Whisper supports 99 languages for speech-to-text. The quality varies by language -- English, Spanish, French, German, Chinese, Japanese, and Korean have the best accuracy. Less common languages may have lower accuracy but are still functional.

The Conjure UI is available in English, German, Spanish, French, Italian, Korean, Portuguese, Brazilian Portuguese, Simplified Chinese, and Traditional Chinese.

What operating systems are supported?

• Windows 10/11 (x64) -- primary platform, fully tested
• Linux (x64) -- Debian/Ubuntu packages and AppImage. Wayland and X11 supported
• macOS -- build support exists but is not the primary target. Metal GPU acceleration is supported

Where is my data stored?

All data is stored locally in a SQLite database (conjure.db) in your system's app data directory:

• Windows: %APPDATA%\com.conjure.desktop\
• Linux: ~/.local/share/com.conjure.desktop/

Transcription history, dictionary entries, preferences, API keys, and conversation data are all in this single file. Audio recordings are stored alongside the database.

Nothing is sent to Conjure's servers. If you use cloud providers, data goes directly from your machine to the provider.

Can I use Conjure with terminals?

Yes. Conjure's streaming text injection (type_text) is specifically designed to work with terminal applications including Windows Terminal, PowerShell, and command-line tools. It uses direct Unicode character injection instead of clipboard paste, which avoids conflicts with terminal key bindings and held modifier keys.

This makes Conjure particularly useful for dictating into AI coding tools (like Claude Code), terminal editors, and command prompts.

How does Conjure compare to Windows Speech Recognition / Voice Typing?

Feature	Conjure	Windows Voice Typing
Offline accuracy	Good (whisper-large-v3)	Good (built-in)
Cloud accuracy	Excellent (Groq, Deepgram)	N/A
Post-processing	Yes (13+ LLM providers)	No
Custom styles	Yes (Polished, Email, Chat, Formal, custom)	No
Voice commands	36 commands	Different set
Dictionary	Yes (glossary + replacements + auto-learn)	Limited
Agent mode	Yes (AI tools, file access, terminal)	No
Open source	Yes (AGPLv3)	No
Terminal support	Yes (streaming injection)	Limited

Dictation

How long can a single dictation be?

With local whisper, dictation is segmented into 300-second (5-minute) chunks that are processed sequentially. There's no hard limit on total recording length.

With cloud providers, limits vary by provider. Most support at least 5-10 minutes per request.

Can I dictate in multiple languages?

Whisper auto-detects the language of your speech. You can switch languages between dictations. However, mixing languages within a single dictation may produce inconsistent results.

Why does my dictation have random text I didn't say?

TIP

This is a known whisper behavior called "hallucination." Increasing the noise gate threshold and speaking promptly after pressing the hotkey are the most effective fixes.

This is whisper hallucination -- a known behavior where the model generates text from silence or noise. Solutions:

1. Increase the noise gate threshold in Settings > Input > Audio
2. Start speaking promptly after pressing the hotkey
3. Use a cloud provider (they have better silence handling)

Can I edit the transcript before it's pasted?

In batch mode, the transcript is pasted immediately after processing. However, you can review and correct past transcriptions from the transcription history page using the inline correction feature.

In the future, a review-before-paste feature may be added.

Technical

What is whisper.cpp?

whisper.cpp is a C/C++ port of OpenAI's Whisper speech recognition model. It runs locally on your machine without needing Python or a GPU. Conjure bundles it as a Rust sidecar using the whisper-rs bindings.

What's the difference between CPU and GPU transcription?

• CPU works on any machine but is slower (5-15 seconds for a 30-second recording)
• GPU (Vulkan on Windows/Linux, Metal on macOS) is 5-10x faster but requires a compatible GPU and drivers

Conjure automatically detects available GPU backends at runtime and falls back to CPU if GPU initialization fails.

What's a sidecar?

The whisper transcription engine runs as a separate process (a "sidecar") alongside the main Conjure application. It's a standalone HTTP server written in Rust that receives audio and returns transcripts. This architecture prevents long transcriptions from blocking the UI.

How does the dictionary-to-transcription feedback loop work?

1. Dictionary glossary entries are concatenated into a text string
2. This string is passed as whisper's initial_prompt parameter
3. Whisper uses the prompt as a vocabulary hint, biasing its recognition toward the words in the prompt
4. Auto-learn corrections that you approve become dictionary entries, which feed back into the next transcription's prompt

This means the more you correct, the more accurate future transcriptions become.

How much disk space does Conjure use?

• App binary: ~50-100MB
• Whisper models: 75MB (tiny) to 3GB (large-v3)
• Database: grows with transcription history (typically under 100MB)
• Audio files: depends on how much you dictate and whether you keep recordings

Can I self-host the cloud backend?

The PocketBase server (server/) is included in the repository and can be self-hosted. However, the cloud mode is not fully operational yet -- handler stubs exist but the transcription proxy, post-processing proxy, and subscription management need implementation.

For now, local-only mode with BYOK API keys provides all the same functionality.

Privacy

Does Conjure collect any data?

No. There is no telemetry, analytics, crash reporting, or usage tracking that reports back to Conjure. All usage statistics (API calls, token counts, word counts) are stored locally for your reference only.

Who can see my dictations?

Only you. Transcriptions are stored locally in SQLite. If you use a cloud provider, your audio is sent to that provider per their privacy policy. Conjure's own servers never see your data.

Are my API keys safe?

API keys are stored in your local SQLite database as plaintext, protected by your operating system's file permissions. They are never transmitted to Conjure's servers -- all API calls go directly from your machine to the provider.

WARNING

If you export your data, API keys are included in the export file. Treat your database file (conjure.db) and export ZIP files as sensitive -- anyone with access can read your keys.

If you export your data, API keys are included in the export file with a warning. Treat your database file and export files as sensitive.

Community

How do I report a bug?

Open an issue on GitHub. Use the bug report template -- it helps us reproduce the issue faster.

How can I contribute?

See CONTRIBUTING.md for setup instructions and contribution guidelines. We welcome bug fixes, documentation, translations, and feature contributions.