verbatim-dicta/README.md

Verbatim Dicta

Real-time audio transcription using Whisper AI with optional LLM-powered analysis. Captures system audio via loopback and transcribes it with configurable models and processing options.

Features

• Real-time transcription of system audio (Windows/Linux)
• Multiple Whisper model sizes (tiny to large)
• Multi-language support
• Sentence extraction mode - Stitches audio chunks into complete sentences
• Optional LLM analysis for fact-checking and question generation (via Ollama)
• GPU acceleration support
• Flexible audio device configuration

Quick Start

# Install dependencies
pip install -r requirements.txt

# Basic transcription (no LLM)
python transcribe_speakers.py

# With LLM analysis (optional)
python transcribe_speakers.py --enable-llm

# With sentence extraction
python transcribe_speakers.py --sentence-mode

# List audio devices
python transcribe_speakers.py --list-devices

Requirements

• OS: Windows 10/11 or Linux
• Python: 3.8+
• Audio: Loopback device (Stereo Mix/VB-Cable on Windows, PulseAudio on Linux)
• Optional: CUDA-capable GPU, Ollama for LLM features

Installation

1. Install Dependencies

pip install -r requirements.txt

2. GPU Support (Optional)

For CUDA 11.8:

pip install torch==2.8.0+cu118 --index-url https://download.pytorch.org/whl/cu118

For CUDA 12.1:

pip install torch==2.8.0+cu121 --index-url https://download.pytorch.org/whl/cu121

3. Audio Loopback Setup

Windows - Option A (Stereo Mix):

1. Right-click speaker icon → Sounds → Recording tab
2. Right-click → Show Disabled Devices
3. Enable and set Stereo Mix as default

Windows - Option B (VB-Cable, recommended):

1. Download from vb-audio.com
2. Install and restart
3. Use --device "CABLE Output"

Linux: Configure PulseAudio loopback or use transcribe_dual_linux.py

4. LLM Features (Optional)

# Install Ollama from ollama.ai
ollama pull llama3.2

Usage

Available Scripts

• transcribe_speakers.py - Main script with all features (LLM optional via --enable-llm)
• transcribe_dual_linux.py - Linux-specific with dual audio support

Common Commands

# Quick start with GPU (English)
./RUN_GPU.sh

# Dutch language
./RUN_DUTCH.sh

# Dutch with LLM analysis
./RUN_DUTCH_LLM.sh

# With LLM analysis
./RUN_GPU.sh --enable-llm

# Save to file
./RUN_GPU.sh --output transcript.txt

# Other languages (Spanish, French, German, etc.)
./RUN_GPU.sh --language es  # Spanish
./RUN_GPU.sh --language fr  # French
./RUN_GPU.sh --language de  # German

# Maximum accuracy with LLM and sentence extraction
python transcribe_speakers.py --model large --enable-llm --sentence-mode --output enriched.txt

# Force CPU (if GPU issues)
python transcribe_speakers.py --force-cpu

Key Options

Option	Description	Default
--model	Model size: tiny/base/small/medium/large	base
--language	Language code (en/es/fr/de/ja/etc.)	en
--device	Audio device name (partial match)	Auto
--interval	Processing interval (seconds)	8.0
--min-duration	Minimum audio duration	3.0
--fast-mode	Fast mode (3-5x faster, lower accuracy)	False
--enable-llm	Enable fact-checking and questions	False
--llm-model	Ollama model to use	llama3.2
--output	Save to file	None
--force-cpu	Disable GPU	False
--gpu-index	GPU device index	0
--sentence-mode	Extract complete sentences from chunks	False

Model Performance

Model	Size	Speed	Quality	Best For
tiny	~75 MB	Fastest	Basic	Quick tests, low-latency
base	~145 MB	Fast	Good	General real-time use
small	~485 MB	Moderate	Better	Balanced accuracy/speed
medium	~1.5 GB	Slow	Great	High accuracy needs
large	~3 GB	Slowest	Best	Maximum accuracy

Optimization Presets

Low Latency (Real-Time):

python transcribe_speakers.py --model tiny --fast-mode --interval 2 --min-duration 1.5

Balanced:

python transcribe_speakers.py --model base --interval 5

High Accuracy:

python transcribe_speakers.py --model large --interval 10 --enable-llm

Troubleshooting

No loopback device:

• Windows: Enable Stereo Mix or install VB-Cable
• Linux: Configure PulseAudio loopback

CUDA errors:

python transcribe_speakers.py --force-cpu

No audio captured:

• Verify audio is playing
• Check device: --list-devices
• Increase system volume

Poor quality:

• Use larger model: --model medium
• Increase interval: --interval 10
• Specify language: --language <code>

Ollama errors:

• Ensure Ollama is running
• Pull model: ollama pull llama3.2

Output Format

Standard:

[14:23:15] Transcribed audio segment.
[14:23:23] Another segment with timestamp.

With LLM (--enable-llm):

======================================================================
[14:23:15] The Earth revolves around the Sun in 365 days.

📊 Fact Check: FACTUAL (confidence: 0.98)
💡 Scientifically accurate. Earth's orbital period is 365.25 days.

❓ Questions:
1. Why do we need leap years?
2. How does Earth's orbit affect seasons?
======================================================================

Technical Stack

• Audio: sounddevice, soundfile (16kHz mono, 16-bit PCM)
• Transcription: faster-whisper (optimized Whisper)
• LLM: Ollama (local inference)
• Capture: WASAPI loopback (Windows), PulseAudio (Linux)

Future Work

• Real-time streaming transcription with reduced buffering
• Speaker diarization improvements
• Web interface for remote monitoring
• Multi-device simultaneous transcription
• Cloud LLM integration options
• Custom vocabulary and domain adaptation
• Noise reduction preprocessing

License

Uses Whisper (OpenAI), faster-whisper (SYSTRAN), and Ollama.