---
library_name: keras
tags:
  - audio-classification
  - cnn
  - cebuano
  - sinama
  - mel-spectrogram
pipeline_tag: audio-classification
---

# Sinama Audio Classifier

A CNN-based audio classification model trained to recognise spoken
Cebuano / Sinama words from short audio clips.

## Usage

### Via Inference API

```python
import requests

API_URL = "https://api-inference.huggingface.co/models/YOUR_USERNAME/sinama-translator"
headers = {"Authorization": "Bearer hf_YOUR_TOKEN"}

with open("audio.wav", "rb") as f:
    response = requests.post(API_URL, headers=headers, data=f.read())

print(response.json())
# [{"label": "ako", "score": 0.95}, ...]
```

### Local inference

```python
import tensorflow as tf, json, librosa, numpy as np

model = tf.keras.models.load_model("best_model.keras")
with open("label_map.json") as f:
    label_map = {int(k): v for k, v in json.load(f).items()}

# preprocess your audio the same way as training …
pred = model.predict(features)
print(label_map[pred.argmax()])
```

## Training details

- **Architecture:** 3-block CNN (Conv2D → BN → ReLU → MaxPool → Dropout)
- **Features:** 128-bin Mel Spectrogram, 4 s clips, 22 050 Hz
- **Optimiser:** Adam
- **Loss:** Categorical cross-entropy