voice-to-vector-database-pipeline

A custom local voice to vector database pipeline.

• OpenAI Whisper transcription
• Qwen 3 embedding model (via Ollama)
• ChromaDB vector database

.
├── src/voice_pipeline/     # Application package
│   ├── config.py           # Shared paths & model settings
│   ├── whisper/            # Audio → text
│   ├── embedding/          # Text → vectors
│   └── vectorstore/        # Vectors → ChromaDB
├── scripts/                # Thin CLI entry points
├── data/                   # Runtime artifacts (gitignored contents)
│   ├── audio/              # pending/ · transcribed/
│   ├── transcripts/        # base/ · pending/ · embedded/
│   └── chroma/             # docs staging + storage/
├── pyproject.toml
└── requirements.txt

• Python 3.10+
• OpenAI Whisper large-v3
• Ollama model qwen3-embedding:8b
• ChromaDB

pip install -e .
# or
pip install -r requirements.txt

1. Place audio files (e.g. .mp3) in data/audio/pending/

2. Transcribe:

   python scripts/transcribe.py
   # or: voice-transcribe

3. Copy a transcribed .txt from data/transcripts/base/<name>/ into data/transcripts/pending/

4. Embed:

   python scripts/embed.py
   # or: voice-embed

5. Store in ChromaDB:

   python scripts/store.py
   # or: voice-store

You should now have a persistent vector database under data/chroma/storage/ with embeddings and documents from the recording.

Transcript and vector filenames change as files move through the pipeline stages. Paths are centralized in src/voice_pipeline/config.py.

Working on automation instead of manually executing each stage.