VOXTERM

Local real-time voice transcription TUI with speaker diarization. Runs entirely offline — no cloud APIs.

Setup

git clone https://github.com/dmarzzz/VoxTerm.git
cd voxterm
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run

python3 app.py

Or use the launcher:

./voxterm

Controls

Key	Action
R	Start/pause recording
T	Tag/name speakers
P	Speaker profiles
M	Switch model
L	Switch language
S	Save transcript
Y	Copy to clipboard
C	Clear transcript
Q	Quit

Voice Tagging

VoxTerm can learn and remember speaker voices across sessions:

1. Record a conversation — speakers are detected as "Speaker 1", "Speaker 2", etc.
2. Press T to name them — type a name, press Enter
3. Next session, VoxTerm auto-recognizes returning speakers with confidence scores
4. The more you tag, the less you need to — the system learns over time

Press P to manage your speaker profile library (rename, delete, wipe all data).

Models

• qwen3-0.6b (default) — fast, ~1.2GB
• qwen3-1.7b — more accurate, ~3.4GB
• Whisper variants (tiny through large-v3) available via M menu

Models download automatically on first use.

Requirements

• macOS with Apple Silicon (M1+)
• Python 3.11+
• Microphone access

Privacy & Voice Data

Voice tagging stores CAM++ speaker embeddings locally on your machine at ~/Library/Application Support/voxterm/.speakers.db. These embeddings are biometric data — they can identify a person across recordings but cannot be used to reconstruct audio.

What we do:

• All processing is local and offline — no data ever leaves your machine
• Embedding BLOBs are encrypted at rest with AES-256-CBC + HMAC-SHA256
• Encryption key is auto-generated and stored in your macOS Keychain — zero config, no passwords
• Existing unencrypted databases are automatically migrated on first open
• Database file permissions are set to owner-only (0600)
• Daily backups with 7-day retention
• Ctrl+X in the profiles screen (P) permanently deletes all voice data with a VACUUM to scrub bytes from disk
• .gitignore excludes all speaker databases and transcripts from version control

Known limitations:

• Metadata is not encrypted. Speaker names, timestamps, and session history are stored in plaintext SQL columns. Only the biometric embedding BLOBs are encrypted. Someone with file access can see who spoke and when, but not derive a voiceprint.
• Keychain access. The encryption key lives in your macOS login Keychain. Any process running as your user could request it. FileVault provides the strongest defense for the at-rest case.

What we explicitly avoid:

• No subprocess calls for key management — the Security framework is called directly via ctypes, so the key never appears in argv or the process list
• No additional dependencies — CommonCrypto and Security.framework are built into macOS