Audio Analysis
Embeddings Vectorizer
Runs each clip through a foundation model — Wav2Vec 2.0 for speech, CLAP for sounds and music — and returns one dense vector (768 or 512 numbers) per clip.
When to use it
Load the vectors into a vector database for similarity search ("find calls like this one"), cluster recordings, or train a small classifier on top.
In a pipeline
Audio File InputAudio Chunk / SlicerEmbeddings VectorizerExport CSV
Settings
| Setting | Type | Required | Default or hint |
|---|---|---|---|
| Model family | select | — | wav2vec2 |
| Model (Hugging Face id) | text | — | facebook/wav2vec2-base |
| Unit length (for cosine search) | checkbox | — | true |
| Run on | select | — | auto |
| Time limit (seconds) | number | — | 900 |
Example
CLAP vectors for sound search
- Model family=CLAP (sound & music, 512-d)
- Unit length (for cosine search)=Ticked
| File Name | Dimensions | Embedding |
|---|---|---|
| glass-01.wav | 512 | [0.0213, -0.0871, …] |
Worth knowing
- Embedding is a JSON array, which most vector databases (pgvector, Qdrant, Pinecone) accept as is.
- Long clips are embedded in pieces and averaged; slice first when each part should get its own vector.
- Models run on the server in FlowDoo's Python and are downloaded once, on first use, into the data folder. A GPU is used when the server has one.