README.md

---
language: en
license: other
tags:
- audio
- bypass-detection
- audio-classification
- content-moderation
- random-forest
- open-source
metrics:
- accuracy
- precision
- recall
- f1
model-index:
- name: open-audiodeleto
  results:
  - task:
      type: audio-classification
    metrics:
    - type: accuracy
      value: 0.70
    - type: precision
      value: 0.6875
    - type: recall
      value: 0.7333
    - type: f1
      value: 0.7097
---

# Open-Audiodeleto

Public/open-source lightweight baseline.

This model detects audio patterns similar to known bypass uploads. It does not prove copyright infringement. It does not prove intent. It should not be used as the sole basis for enforcement.

## What It Does

This model produces a bypass-audio probability from audio-only features such as duration, RMS statistics, spectral statistics, tempo, silence ratio, and loudness-transition scores.

## What It Does Not Do

It does not prove copyright infringement. It does not prove intent. It should not be used for automatic bans, automatic account termination, or legal copyright conclusions. No automatic enforcement should rely only on this model.

## Run

```bash
python open_audiodeleto.py /path/to/audio.ogg
```

Example:

```bash
python open_audiodeleto.py audio.mp3
```

Example JSON:

```json
{
  "model": "Open-Audiodeleto",
  "version": "0.1-800",
  "file": "audio.mp3",
  "bypassProbability": 0.87,
  "riskLevel": "high",
  "recommendation": "review",
  "note": "This model detects audio patterns similar to known bypass uploads. It does not prove copyright infringement. It does not prove intent. It should not be used as the sole basis for enforcement."
}
```

## Metrics

Validation:

```json
{
  "accuracy": 0.7,
  "precision": 0.6875,
  "recall": 0.7333333333333333,
  "f1": 0.7096774193548387,
  "confusion_matrix": [
    [
      40,
      20
    ],
    [
      16,
      44
    ]
  ]
}
```

Test:

```json
{
  "accuracy": 0.7916666666666666,
  "precision": 0.7692307692307693,
  "recall": 0.8333333333333334,
  "f1": 0.8,
  "confusion_matrix": [
    [
      45,
      15
    ],
    [
      10,
      50
    ]
  ]
}
```

## Files

- `model.pkl`
- `metadata.json`
- `features.csv`
- `splits.csv`
- `results.json`
- `misclassified_files.csv`
- `invalid_files.csv`
- `open_audiodeleto.py`
- `requirements.txt`

<!-- ONNX_EXPORT_SECTION_START -->
## Purpose

Public lightweight baseline audio-bypass risk classifier.

Audiodeleto is an audio-bypass risk signal only. It does not prove copyright infringement, intent, or a policy violation, and should not be the sole basis for enforcement.

## Model Files

- `bin/audiodeleto.py`: command-line scorer.
- `model/model.pkl`: Python/scikit-learn model artifact.
- `model/model.onnx`: cross-platform ONNX model artifact.
- `model/metadata.json`: thresholds, version, feature configuration, calibration metadata.
- `model/feature_spec.json`: canonical feature names and order.
- `reports/`: validation, ONNX export, and misclassification reports.
- `training/`: feature rows and split files used to package this release.

## Runtime Artifacts

### Python usage (`model.pkl`)

Use `model/model.pkl` with the exact feature order in `model/feature_spec.json` when running inside Python:

```python
import json
import joblib
import pandas as pd

model = joblib.load("model/model.pkl")
feature_spec = json.load(open("model/feature_spec.json", encoding="utf-8"))
features = pd.DataFrame([values], columns=feature_spec["featureNames"])
probability = float(model.predict_proba(features)[0, 1])
```

### Cross-platform usage (`model.onnx`)

Use `model/model.onnx` from any ONNX Runtime host. The ONNX graph expects precomputed audio features, not raw audio.

- Model ID: `open-audiodeleto`
- Version: `0.1-800`
- Input name: `features`
- Input type: `float32`
- Input shape: `[batch, 15]`
- Feature order: `feature_spec.json`
- Probability output: `probabilities[:, 1]`

### Feature spec warning

Do not infer feature order from CSV column order or object key order. Always use `model/feature_spec.json`; a wrong feature order can produce valid-looking but incorrect probabilities.

<!-- ONNX_EXPORT_SECTION_END -->