Private, local-only desktop app for audio transcription with speaker diarization and AI-powered summarization.

Diaricat is a Windows desktop application that transcribes audio/video files, identifies who said what (speaker diarization), and generates AI-powered summaries — all running locally on your machine. No data ever leaves your computer.

Diaricat is part of a broader vision for local-first AI systems focused on privacy, autonomy, and offline intelligence.

• Accurate transcription powered by Faster Whisper (large-v3 model with CUDA acceleration)
• Speaker diarization using SpeechBrain ECAPA-TDNN embeddings with custom agglomerative clustering
• AI correction & summarization with local LLM inference via llama.cpp, running GGUF models such as Qwen 2.5 7B Instruct (Q4_K_M)
• 100% offline & private — no API keys, no cloud services, no data upload
• Bilingual UI — Spanish and English with one-click toggle
• Multiple export formats — TXT, SRT, DOCX, PDF, JSON
• Modern dark UI — glassmorphism design built with React + Tailwind CSS

Diaricat follows a design language I call Purple Space Glass.

It blends glassmorphism, deep-space aesthetics and soft neon reflections to create interfaces that feel both modern and fluid — almost like interacting with an intelligent system rather than a static tool.

The goal is not just visual appeal, but to make AI systems feel:

• responsive
• ambient
• alive, without being intrusive

This design direction is part of a broader vision where local AI systems are not only powerful and private, but also intuitive and pleasant to use.

+---------------------------------------------------+
|                  Desktop Shell                     |
|              (pywebview + .NET/Edge)               |
+---------------------------------------------------+
|                  Frontend (UI)                     |
|        React  TypeScript  Vite  Tailwind           |
|           Radix UI  Lucide  shadcn/ui              |
+---------------------------------------------------+
|                 REST API Layer                     |
|           FastAPI  Uvicorn  Pydantic               |
+--------------+------------+-----------+-----------+
| Transcription|Diarization | LLM Post- |  Export   |
|   Service    |  Service   | process   |  Service  |
|  (Whisper)   |(SpeechBrain)|(llama.cpp)|(DOCX/PDF) |
+--------------+------------+-----------+-----------+
|               Pipeline Orchestrator                |
|        Job queue  Progress  Cancellation           |
+---------------------------------------------------+

• Python 3.11+ (tested with 3.14)
• Node.js 18+ (for frontend)
• NVIDIA CUDA Toolkit 12.x (for GPU acceleration)
• Visual Studio Build Tools 2022 (for building llama-cpp-python)

1. Download the latest release from Releases
2. Extract the Diarcat/ folder
3. (Optional) Place a GGUF model in workspace/models/ for AI summaries
4. Run Diarcat.exe

# Clone the repository
git clone https://github.com/nia-huck/Diaricat.git
cd Diaricat

# Create Python virtual environment
python -m venv .venv
.venv\Scripts\activate

# Install Python dependencies
pip install -e ".[dev]"

# Install torch with CUDA (optional, for GPU acceleration)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu128

# Install faster-whisper and speechbrain
pip install faster-whisper speechbrain

# Install llama-cpp-python (requires Visual Studio Build Tools)
pip install llama-cpp-python

# Install frontend dependencies
cd frontend
npm install

# Start frontend dev server
npm run dev

# In another terminal, start the backend
cd ..
python -m diaricat api --host 127.0.0.1 --port 8765

# Build the frontend
cd frontend
npm run build
cd ..

# Run PyInstaller
python -m PyInstaller packaging/diaricat.spec --distpath dist --workpath build --noconfirm

# Output: dist/Diarcat/Diarcat.exe

The build uses onedir mode for fast startup (~1 second vs minutes for onefile).

Diaricat uses llama.cpp as the local inference engine for transcript correction and summarization, running GGUF-compatible models. Without a configured model, it falls back to rule-based processing.

• Qwen 2.5 7B Instruct (Q4_K_M) — configured by default for the local post-processing pipeline

• Default model path: models/qwen2.5-7b-instruct-q4_k_m.gguf
• Context size: 4096
• Inference engine: llama.cpp

Place the .gguf file in workspace/models/ and configure the path in Settings. Model path and runtime parameters can be updated from the application Settings screen.

1. Validation — Verify source file exists and is a supported format
2. Audio normalization — Extract audio, convert to 16kHz mono WAV via FFmpeg
3. Transcription — Speech-to-text with Faster Whisper (chunked for long audio)
4. Speaker diarization — Identify speakers using ECAPA-TDNN embeddings + agglomerative clustering
5. Segment merge — Align ASR segments with speaker turns
6. Correction (optional) — Fix ASR errors using local LLM
7. Summarization (optional) — Generate structured summary with key points, decisions, and topics

Diaricat/
├── src/diaricat/           # Python backend
│   ├── api/                # FastAPI routes and middleware
│   ├── core/               # Orchestrator, job queue, alignment
│   ├── services/           # Transcription, diarization, postprocess, export
│   ├── models/             # Pydantic domain and API models
│   ├── utils/              # Device detection, logging, compatibility
│   ├── desktop.py          # pywebview desktop shell
│   ├── main.py             # CLI entry point
│   └── settings.py         # Configuration management
├── frontend/               # React/TypeScript UI
│   └── src/
│       ├── components/     # UI components (screens, ui primitives)
│       ├── context/        # React context (AppContext, I18nContext)
│       ├── lib/            # API client, i18n translations
│       └── types/          # TypeScript type definitions
├── config/                 # Default configuration (YAML)
├── packaging/              # PyInstaller spec and runtime hooks
├── scripts/                # Build scripts
├── tests/                  # Unit tests
├── vendor/                 # Bundled FFmpeg binaries
└── pyproject.toml          # Project metadata and dependencies

Settings are stored in config/default.yaml and can be modified through the Settings screen in the app:

Backend: Python 3.14 · FastAPI · Uvicorn · Pydantic · PyYAML · SpeechBrain · Faster Whisper · CTranslate2 · llama-cpp-python · PyInstaller

Frontend: React 18 · TypeScript · Vite · Tailwind CSS · Radix UI · shadcn/ui · Lucide Icons

AI Models: Whisper large-v3 (ASR) · pyannote/speaker-diarization-3.1 + ECAPA-TDNN embeddings (diarization) · Qwen 2.5 7B Instruct Q4_K_M (correction/summary via llama.cpp)

MIT License — see LICENSE file for details.

Built with privacy in mind. Your audio never leaves your machine.