README.md

---
license: apache-2.0
base_model:
- lj1995/VoiceConversionWebUI
- facebook/hubert-base-ls960
pipeline_tag: feature-extraction
library_name: fairseq
tags:
- rvc
- audio
---

# Hubert Base ONNX Model for Voice Conversion

This is the **ONNX-exported version of the Hubert Base model**, fine-tuned for voice conversion and compatible with modern inference pipelines. This model allows fast and efficient audio processing in ONNX runtime environments.

It builds upon the following models:
- [lj1995/VoiceConversionWebUI](https://huggingface.co/lj1995/VoiceConversionWebUI)
- [facebook/hubert-base-ls960](https://huggingface.co/facebook/hubert-base-ls960)

---

## Features

- Converts audio features into high-quality embeddings for voice conversion tasks.
- Fully ONNX-compatible for optimized inference on CPUs and GPUs.
- Lightweight and easy to integrate in custom voice processing pipelines.
- No extra requirements needed, just **numpy** and **onnxruntime**

## ONNX Model Report

**Model:** `hubert_base.onnx`  
**Producer:** pytorch 2.0.0  
**IR Version:** 8  
**Opsets:** ai.onnx:18  
**Parameters:** 94,370,816  

---

### 🟦 Inputs
- **source** | type: `float32` | shape: [batch_size, sequence_length]
  - *Waveform PCM 32 - SR 16,000 - Mono*
- **padding_mask** | type: `bool` | shape: [batch_size, sequence_length]
  - It is usually a completely false array, with the same shape as the waveform. `padding_mask = np.zeros(waveform.shape, dtype=np.bool_)`

### 🟩 Outputs
- **features** | type: `float32` | shape: [batch_size, sequence_length, 768 ]



---

## Usage

```python
import numpy as np
import onnxruntime as ort

class OnnxHubert:
    """
    Class to load and run the ONNX model exported by Hubert.
    
    Attributes:
        session (ort.InferenceSession): The ONNX Runtime session.
        input_name (str): The name of the input node.
        output_name (str): The name of the output node.
    
    Methods:
        extract_features_batch (source, padding_mask): Run the ONNX model and extract features from the batch.
        extract_features (source, padding_mask): Run the ONNX model and extract features from a single input.
    """
    def __init__(self, model_path: str, thread_num: int = None):
        """
        Initialize the OnnxHubert object.

        Parameters:
            model_path (str): The path to the ONNX model file.
            thread_num (int, optional): The number of threads to use for inference. Defaults to None.

        Attributes:
            session (ort.InferenceSession): The ONNX Runtime session.
            input_name (str): The name of the input node.
            output_name (str): The name of the output node.
        """
        self.session = ort.InferenceSession(model_path)
        
        self.input_name = self.session.get_inputs()[0].name
        self.output_name = self.session.get_outputs()[0].name
    def extract_features(
        self, 
        source: np.ndarray, 
        padding_mask: np.ndarray
    ) -> np.ndarray:
        """
        Extract features from the batch using the ONNX model.

        Inputs:
            source: ndarray of shape (batch_size, sequence_length) float32
            padding_mask: ndarray of shape (batch_size, sequence_length) bool

        Returns:
            ndarray of shape (D, 768) with the extracted features
        """
        result = self.session.run(None, {
            "source": source,
            "padding_mask": padding_mask
        })
        return result[0]
```

## Installation

You can install the required libraries with:

```bash
pip install onnxruntime numpy
```