Cognition

Cognition v1 Features

• Meeting Agent: Does exactly what Granola does – captures everything you hear, no bots needed. Transforms messy meeting notes into clean, actionable and shareable action items and MOMs.

• Sales Agent: Real-time contextual intel like Cluely. It will literally tell you "your prospect mentioned budget concerns at the 15th minute – you missed probing that." and all the phrasing is backed by SPIN, MEDDIC selling techniques. Not generic advice – specific missed opportunities called out in real-time. Saves you from that post-meeting "damn, should have said this" moment.

• Transparent and Invisible: Yes – it won't show up in screen recording, screen shares and stays persistent across macOS multiple desktops and apps.

---

▶️ Watch Cognition Demo on Loom

Features

• 🚀 Real-time transcription from microphone input
• 📊 Comprehensive benchmarking of different models and configurations
• 🎯 3-second chunk processing for near real-time results
• 🔧 Multiple model sizes (tiny, base, small, medium, large)
• 💻 CPU and GPU support with optimized compute types
• 🎤 VAD (Voice Activity Detection) filtering for better accuracy

Quick Start

1. Install Dependencies

# Install Python dependencies
pip install -r requirements.txt

# On macOS, you might need to install portaudio for audio capture
brew install portaudio

2. Run Real-time Transcription Test

python realtime_transcription_test.py

The script will:

• Let you choose model size (tiny, base, small, medium, large)
• Ask if you want to use GPU (if available)
• Start capturing audio from your microphone
• Transcribe 3-second chunks in real-time
• Display results with processing time

3. Run Benchmark Tests

python benchmark_test.py

This will test all model configurations and show:

• Load times for each model
• Transcription speed (Real-time Factor)
• Best configurations for speed vs accuracy

Model Performance Guide

Model Size	Speed	Accuracy	Memory	Best For
tiny	⚡⚡⚡⚡⚡	⭐⭐	39MB	Fastest real-time
base	⚡⚡⚡⚡	⭐⭐⭐	74MB	Good balance
small	⚡⚡⚡	⭐⭐⭐⭐	244MB	Better accuracy
medium	⚡⚡	⭐⭐⭐⭐⭐	769MB	High accuracy
large	⚡	⭐⭐⭐⭐⭐	1550MB	Highest accuracy

Real-time Performance

CPU Performance (Intel/AMD)

• tiny model: ~0.3x real-time (very fast)
• base model: ~0.8x real-time (near real-time)
• small model: ~2.0x real-time (slower)
• medium model: ~5.0x real-time (much slower)

GPU Performance (NVIDIA)

• tiny model: ~0.1x real-time (extremely fast)
• base model: ~0.3x real-time (very fast)
• small model: ~0.8x real-time (near real-time)
• medium model: ~1.5x real-time (faster than real-time)

Configuration Options

Model Sizes

• tiny: Fastest, lowest accuracy, good for real-time
• base: Good balance of speed and accuracy
• small: Better accuracy, still reasonably fast
• medium: High accuracy, slower
• large: Highest accuracy, slowest

Compute Types

• int8: Quantized, fastest, works on CPU
• float16: Half precision, good for GPU
• float32: Full precision, slowest

Devices

• cpu: Works everywhere, slower
• cuda: GPU acceleration, much faster (requires NVIDIA GPU)

Usage Examples

Basic Real-time Transcription

from realtime_transcription_test import RealtimeTranscriber

# Initialize with base model on CPU
transcriber = RealtimeTranscriber(
    model_size="base",
    device="cpu",
    compute_type="int8"
)

# Start transcription
transcriber.start_transcription()

GPU Acceleration

# Use GPU for faster processing
transcriber = RealtimeTranscriber(
    model_size="small",
    device="cuda",
    compute_type="float16"
)

Custom Chunk Duration

# Modify the chunk duration in the class
transcriber.chunk_duration = 5.0  # 5 seconds instead of 3
transcriber.chunk_size = int(transcriber.sample_rate * transcriber.chunk_duration)

System Requirements

Minimum Requirements

• Python 3.8+
• 4GB RAM
• Microphone access

Recommended for Real-time

• Python 3.9+
• 8GB+ RAM
• NVIDIA GPU (for CUDA acceleration)
• SSD storage

macOS Specific

# Install audio dependencies
brew install portaudio

# If you get audio permission errors, grant microphone access in System Preferences

Windows Specific

# Install Visual C++ Build Tools if you get compilation errors
# Download from: https://visualstudio.microsoft.com/visual-cpp-build-tools/

Troubleshooting

Audio Permission Issues

• macOS: Go to System Preferences > Security & Privacy > Microphone
• Windows: Check microphone permissions in Settings > Privacy > Microphone
• Linux: Ensure your user is in the audio group

CUDA Issues

# Check if CUDA is available
python -c "import torch; print(torch.cuda.is_available())"

# If not available, install CUDA toolkit or use CPU

Memory Issues

• Use smaller models (tiny, base) for limited RAM
• Close other applications
• Use int8 quantization on CPU

Slow Performance

• Use GPU if available
• Use smaller models
• Reduce chunk duration
• Use int8 quantization

Advanced Usage

Custom Audio Sources

# Modify the audio_callback method to use different audio sources
def custom_audio_callback(self, indata, frames, time, status):
    # Process audio from different sources
    # e.g., system audio, specific applications, etc.
    pass

Batch Processing

from faster_whisper import WhisperModel, BatchedInferencePipeline

model = WhisperModel("base", device="cuda", compute_type="float16")
batched_model = BatchedInferencePipeline(model=model)

# Process multiple audio files
segments, info = batched_model.transcribe("audio.mp3", batch_size=16)

Word-level Timestamps

segments, _ = model.transcribe("audio.mp3", word_timestamps=True)

for segment in segments:
    for word in segment.words:
        print(f"[{word.start:.2f}s -> {word.end:.2f}s] {word.word}")

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• faster-whisper by SYSTRAN
• OpenAI Whisper for the original model
• CTranslate2 for fast inference