---
license: mit
library_name: pytorch
tags:
- audio-classification
- speech
- music
- pytorch
pipeline_tag: audio-classification
---

# SMAD Audio Classifier

Classifies audio into four labels:

| id | label | meaning |
|---:|---|---|
| 0 | `speech_noise` | spoken voice over non-music noise |
| 1 | `speech_music` | spoken voice over a music bed |
| 2 | `singing_music` | sung voice over music |
| 3 | `none` | no human voice: instrumental music, noise, or silence |

## Install

```bash
pip install torch librosa soundfile numpy
```

## Use

```python
from predict_crnn import SmadClassifier

clf = SmadClassifier()
result = clf.predict("audio.mp3")

print(result["overall"])
print(result["share"])
```

CLI:

```bash
python predict_crnn.py audio.mp3 --json
```

Each 4-second window includes a label, confidence, and per-class probabilities.
The `overall` label is the class covering the most windows in the file.