QazMeeting AI

On-premise meeting assistant for Russian, Kazakh, and mixed-language meetings. The existing dashboard, demo meeting, SQLite storage, meeting view, and DOCX/PDF exports remain available. Runtime audio and transcripts stay on the local machine; there is no OpenAI, cloud STT, or cloud LLM integration.

Demo mode (zero model dependencies)

From the repository directory in Windows PowerShell:

py -3 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
$env:DEMO_MODE = "true"
python -m uvicorn app.main:app --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000. Demo mode is the default. It uses the existing Russian/Kazakh/mixed-language sample meeting and does not import or need faster-whisper or download any AI models. New meeting creation and exports continue to work as before. Demo data is stored locally in data/qazmeeting.db.

Real local speech-to-text

Audio upload is enabled when DEMO_MODE=false. Install faster-whisper in the virtual environment; it is deliberately excluded from the lightweight demo requirements:

.\.venv\Scripts\Activate.ps1
python -m pip install faster-whisper
$env:DEMO_MODE = "false"
$env:WHISPER_MODEL = "small"
$env:WHISPER_DEVICE = "cpu"
$env:WHISPER_COMPUTE_TYPE = "int8"
python -m uvicorn app.main:app --host 127.0.0.1 --port 8000

The first model initialization requires the selected CTranslate2 model to be available in the local cache. faster-whisper may download the model weights on first use; prepare/cache the model while connected if the runtime machine is offline. No model download occurs during the demo setup or automated checks. The suggested development model is small; WHISPER_MODEL also accepts faster-whisper model names or a local model directory. The default language is automatic (language=None) so Russian, Kazakh, and code-switching are not forced into one language. Segment timestamps and detected language are saved with the recognized text.

Environment variables:

Variable	Default	Purpose
DEMO_MODE	true	Select demo adapters or local transcription/upload
WHISPER_MODEL	small	faster-whisper model name or local model path
WHISPER_DEVICE	cpu	Inference device, such as cpu or cuda
WHISPER_COMPUTE_TYPE	int8	CTranslate2 compute type; int8 is suitable for CPU
DATABASE_PATH	data/qazmeeting.db	Local SQLite database path
UPLOAD_DIR	data/uploads	Private local audio upload directory

For GPU, install a compatible CUDA/cuDNN and CTranslate2 setup for your hardware, then set WHISPER_DEVICE=cuda and a supported compute type such as float16. CPU is the supported simple setup; GPU setup depends on the local driver/runtime.

Supported uploads: .mp3, .wav, .m4a, .mp4 (maximum 1 GB). In local-AI mode click New meeting / Upload recording, enter a title, choose the recording, and select Upload and transcribe locally. The page reports progress during processing and opens the saved transcript when complete. Invalid formats and processing failures are reported in the UI.

Privacy and storage

Inference runs through faster-whisper on this machine. Uploaded recordings are written to data/uploads; meeting text and timestamps are stored in the local SQLite database. These runtime files, .env, exports, and common model/cache directories are excluded from Git. Keep the host filesystem and any backups under your organization’s control. Avoid serving the app on an untrusted network; the current hackathon app does not implement user authentication.

Current limitations

• Local transcription supports Russian and Kazakh through the multilingual Whisper model; mixed-language recognition quality depends on the model and audio.
• A transcript segment currently receives the temporary label Speaker 1 (unverified). It is not a diarization result.
• Real speaker diarization is the NEXT implementation step. The transcription, diarization, and aligned-transcript domain structures are separated so a future local pyannote adapter can align speaker turns with Whisper segments.
• Summaries and action items are not generated from uploaded recordings in this iteration. The demo meeting retains its existing sample summary and action items.
• No real transcription confidence values are fabricated.

Development checks

python -m compileall app

This check does not install dependencies or download model files.