File size: 4,433 Bytes
8732fc1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
# Vibe Search for AudioMuse-AI

Search your own music library by how songs sound and feel ("rainy Sunday coffee shop", "upbeat but melancholic, no vocals", "something for a long night drive"). You can also find songs with the same vibe as one you pick, and save any result list as a playlist on your media server.

The model is small and runs on the CPU. It sits next to AudioMuse-AI's own analysis and leaves it alone: the plugin keeps its vectors in its own table and never edits AudioMuse-AI's data.

## What it does

- **Vibe vectors for every song.** Each time AudioMuse-AI analyses a song, the plugin runs a small audio model (EfficientAT mn20_as, distilled from a combined LCO-Embedding-Omni-7B + e5-omni-7B teacher) over the song in 10-second windows. It stores one 512-number vector per song.
- **Index my library.** This fills in vectors for songs that were analysed before the plugin was installed. It runs as a background task and reports its progress.
- **Search.** Your search text goes through the matching text model (Qwen3-Embedding-0.6B plus a small trained layer) and is compared with every song's vector.
- **Same vibe.** Pick a song to list the songs whose vectors are closest to it.
- **Playlists.** Save a result list to your media server in one click.

Model version: v1.2. Benchmarks and limits are on the model card: https://huggingface.co/Divi19/vibe-embed

## Install

Requires AudioMuse-AI 3.7.1 or later.

1. In AudioMuse-AI, open **Plugins**, add this repository URL, and refresh the catalog:
   `https://huggingface.co/Divi19/vibe-embed/resolve/main/plugin/manifest.json`
2. Install **Vibe Search**, then apply. AudioMuse-AI restarts its workers.
3. Open the Vibe Search page and press **Index my library** once. From then on, new songs get their vectors automatically when they are analysed.

On first use, the plugin downloads its models into `PLUGINS_DIR/_data/vibe_search/models/`: about 35 MB for the audio model and about 818 MB for the 8-bit search model. It needs no extra Python packages: it uses the ONNX Runtime, numpy and tokenizers that AudioMuse-AI already ships with.

### ONNX Runtime versions

- The 8-bit search model needs onnxruntime 1.28 or later, which the standard AudioMuse-AI images have.
- On older onnxruntime builds (the no-AVX2 image ships 1.19), the plugin uses the full-precision search model instead (about 2396 MB, more RAM). It does this by itself and says so on its page. Results are the same.
- If an 8-bit file fails to load for any other reason, the plugin falls back to full precision the same way.

## Settings

| setting | default | meaning |
|---|---|---|
| model_repo | `Divi19/vibe-embed` | Hugging Face repository with the model files |
| revision | `main` | branch, tag or commit of that repository |
| base_url | `https://huggingface.co` | change this only for a mirror |
| precision | `int8` | `int8` (smaller, faster) or `fp32` |
| threads | 4 | CPU threads for the models |
| embed_on_analysis | on | embed each song right after AudioMuse-AI analyses it |
| result_limit | 50 | songs shown per search |

If you change model_repo or revision, press **Index my library** again. Vectors from a different model are not mixed in: every vector is stored with the tag of the model that made it. Changing precision does not need re-indexing.

## Safety checks

The model repository includes `parity.json`, which holds reference outputs recorded when the models were exported. Before first use, the plugin runs a test signal and five test searches through the downloaded models and compares the results (and the search tokens) with those references. If anything does not match (a wrong file, a broken download or a tokenizer mismatch), it refuses to run and shows the reason on its page. It does not silently return bad results.

## Model files

The plugin expects this layout in the Hugging Face repository:

```
config.json
parity.json
mn20_as/audio_encoder_int8.onnx      mn20_as/audio_encoder.onnx
text/text_encoder_int8.onnx(.data)   text/text_encoder.onnx(.data)
text/tokenizer/tokenizer.json        text/tokenizer/tokenizer_config.json
```

## Licences

- The plugin code is licensed under AGPL-3.0 (see LICENSE), the same as AudioMuse-AI, because it runs inside AudioMuse-AI.
- The models it downloads are released separately under Apache-2.0.
- The training music is CC0 / CC-BY / public-domain audio; the attribution list is in the model repository.