### Install VoiceFixer from Source

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Install the latest development version of VoiceFixer directly from its GitHub repository using pip. This is an alternative if you do not trust PyPI or want the bleeding-edge version.

```bash
pip install git+https://github.com/fakerybakery/voicefixer
```

--------------------------------

### Install FFmpeg on Windows

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Install FFmpeg on Windows using the Scoop package manager. This command ensures FFmpeg is available for processing non-WAV audio files.

```bash
scoop install main/ffmpeg
```

--------------------------------

### Install VoiceFixer 2

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Install VoiceFixer 2 using pip. For non-WAV file support, FFmpeg must also be installed.

```bash
pip install voicefixer2
```

```bash
# macOS
brew install ffmpeg

# Ubuntu/Debian
sudo apt install ffmpeg

# Windows (scoop)
scoop install main/ffmpeg
```

```bash
pip install git+https://github.com/voicefixer/voicefixer
```

--------------------------------

### Install FFmpeg on Linux/Ubuntu

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Install FFmpeg on Linux distributions like Ubuntu using the apt package manager. FFmpeg is necessary for handling various audio formats beyond WAV.

```bash
sudo apt install ffmpeg
```

--------------------------------

### Include VoiceFixer in setup.py (PyPI)

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Add voicefixer2 to your setup.py file for installation from PyPI. This is the standard way to list package dependencies in a Python project.

```python
[
    'voicefixer2',
]
```

--------------------------------

### Install VoiceFixer2 via PyPI

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Install the latest published release of the voicefixer2 package using pip. This is the standard method for installing Python packages.

```bash
pip install voicefixer2
```

--------------------------------

### Install FFmpeg on macOS

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Install FFmpeg on macOS using the Homebrew package manager. FFmpeg is required for non-WAV audio file support.

```bash
brew install ffmpeg
```

--------------------------------

### Include VoiceFixer in setup.py (Git URL)

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Add voicefixer2 to your setup.py file using a direct Git URL. This method is common for packages installed from version control systems.

```python
[
    'voicefixer2 @ git+https://github.com/voicefixer/voicefixer',
]
```

--------------------------------

### Include VoiceFixer2 in requirements.txt

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Add voicefixer2 to your project's requirements.txt file for easy dependency management. This ensures the correct version is installed when setting up the project.

```text
voicefixer2
```

--------------------------------

### Get VoiceFixer CLI help

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Display the help message for the VoiceFixer command-line interface. This provides information on all available options and their usage.

```bash
voicefixer -h
```

--------------------------------

### Process all audio files in a directory using CLI

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Batch process all audio files within a specified input folder and save the results to an output folder using the VoiceFixer CLI. Ensure FFmpeg is installed for non-WAV files.

```bash
voicefixer --infolder /path/to/input --outfolder /path/to/output
```

--------------------------------

### Include VoiceFixer from Git in requirements.txt

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Specify the voicefixer package to be installed directly from its GitHub repository within your requirements.txt file. This is useful for development versions or specific forks.

```text
git+https://github.com/voicefixer/voicefixer
```

--------------------------------

### Restore Audio In-Memory with VoiceFixer

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Load an audio file, restore it in memory using VoiceFixer, and save the result. Ensure librosa and soundfile are installed for loading and saving audio.

```python
import librosa
import soundfile as sf
from voicefixer import VoiceFixer

# Load audio manually at 44100 Hz
wav, _ = librosa.load("degraded_speech.wav", sr=44100)

# Restore in-memory (returns np.ndarray)
restored_wav = voicefixer.restore_inmem(
    wav_10k=wav,
    cuda=False,  # or True for GPU
    mode=0
)

# Save result using soundfile
sf.write("restored_output.wav", restored_wav, samplerate=44100)
print(f"Restored audio shape: {restored_wav.shape}")
```

--------------------------------

### VoiceFixer.__init__

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Initializes the VoiceFixer restoration model, downloading weights from Hugging Face. The `model` parameter accepts any Hugging Face model repo ID containing a `vf.ckpt` checkpoint file.

```APIDOC
## VoiceFixer.__init__(model='voicefixer/voicefixer')

### Description
Initialize the VoiceFixer restoration model, downloading weights from Hugging Face

`VoiceFixer` is a `torch.nn.Module` that loads a pretrained speech restoration checkpoint from the Hugging Face Hub. The `model` parameter accepts any Hugging Face model repo ID that contains a `vf.ckpt` checkpoint file. On first use, the checkpoint is downloaded and cached locally via `cached_path`. The model operates at 44.1 kHz and supports 2-channel (stereo) output.

### Parameters
#### Path Parameters
- **model** (string) - Optional - Hugging Face model repo ID containing a `vf.ckpt` checkpoint file. Defaults to 'voicefixer/voicefixer'.

### Request Example
```python
from voicefixer import VoiceFixer

# Load with default pretrained model (voicefixer/voicefixer on Hugging Face)
voicefixer = VoiceFixer()

# Load with a custom Hugging Face model repo
voicefixer_custom = VoiceFixer(model='your-org/your-voicefixer-model')
# Raises RuntimeError if vf.ckpt is not found in the cache or repo
```
```

--------------------------------

### Command Line - Help

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Display help information for the VoiceFixer command-line interface.

```APIDOC
## voicefixer -h

### Description
Displays detailed help information for the VoiceFixer command-line tool, including all available options and their descriptions.

### Usage
```bash
voicefixer -h
```
```

--------------------------------

### Initialize VoiceFixer Model

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Initialize the VoiceFixer restoration model, optionally loading a custom model from Hugging Face. The checkpoint is downloaded and cached locally on first use.

```python
from voicefixer import VoiceFixer

# Load with default pretrained model (voicefixer/voicefixer on Hugging Face)
voicefixer = VoiceFixer()

# Load with a custom Hugging Face model repo
voicefixer_custom = VoiceFixer(model='your-org/your-voicefixer-model')
# Raises RuntimeError if vf.ckpt is not found in the cache or repo
```

--------------------------------

### Python API - Initialization

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Initialize the VoiceFixer class for use in Python scripts.

```APIDOC
## VoiceFixer()

### Description
Initializes the VoiceFixer object for programmatic use.

### Usage
```python
from voicefixer import VoiceFixer

# Initialize with default model
voicefixer = VoiceFixer()

# Initialize with a specific model (optional)
# voicefixer = VoiceFixer(model='voicefixer/voicefixer')
```

### Parameters
#### Initialization Parameters
- **model** (string) - Optional - Specifies the model to use. Defaults to the original model. Example: `'voicefixer/voicefixer'`.
```

--------------------------------

### Initialize and Run Standalone Neural Vocoder

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Synthesize speech waveforms from mel spectrograms using the standalone `Vocoder`. It expects non-normalized mel spectrograms and supports both CPU and GPU inference.

```python
import torch
import librosa
import numpy as np
from voicefixer import Vocoder

# Initialize vocoder at 44.1 kHz
vocoder = Vocoder(sample_rate=44100)

# Create a dummy mel spectrogram (batch=1, channels=1, time=100, mel_bins=128)
mel = torch.rand(1, 1, 100, 128)  # replace with real mel spectrogram

# Synthesize waveform
with torch.no_grad():
    waveform = vocoder(mel, cuda=False)   # returns [1, 1, num_samples]

print(f"Output waveform shape: {waveform.shape}")

# GPU inference on CUDA
if torch.cuda.is_available():
    waveform_gpu = vocoder(mel, cuda=True)
```

--------------------------------

### Command Line - Process Directory

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Process all audio files within a specified directory using the VoiceFixer command-line interface.

```APIDOC
## voicefixer --infolder --outfolder

### Description
Processes all audio files in a given input folder and saves them to an output folder.

### Usage
```bash
voicefixer --infolder <path_to_input_folder> --outfolder <path_to_output_folder>
```

### Parameters
#### Command Line Arguments
- **--infolder** (string) - Required - Path to the directory containing input audio files.
- **--outfolder** (string) - Required - Path to the directory where processed audio files will be saved.
```

--------------------------------

### Run all VoiceFixer modes via CLI

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Execute the VoiceFixer CLI with the 'all' mode to process an audio file through all available restoration modes. This is useful for comparing results across different settings.

```bash
voicefixer --infile /path/to/input.wav --outfile /path/to/output.wav --mode all
```

--------------------------------

### Command Line - Process Single File

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Process a single audio file using the VoiceFixer command-line interface.

```APIDOC
## voicefixer --infile

### Description
Processes a single audio file.

### Usage
```bash
voicefixer --infile <path_to_input_file>
```

### Parameters
#### Command Line Arguments
- **--infile** (string) - Required - Path to the input audio file (e.g., `test/utterance/original/original.wav`).
- **--outfile** (string) - Optional - Path to save the processed audio file. Defaults to `outfile.wav` if not specified.
- **--mode** (integer or 'all') - Optional - Specifies the processing mode. Defaults to 0. Can be an integer for specific modes or 'all' to run all modes.
```

--------------------------------

### Initialize VoiceFixer Python API

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Instantiate the VoiceFixer class in Python to use its audio restoration capabilities programmatically. You can optionally specify a model name.

```python
from voicefixer import VoiceFixer
voicefixer = VoiceFixer()
# or voicefixer = VoiceFixer(model='voicefixer/voicefixer')
```

--------------------------------

### Public API - Process Audio

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Use the free public API for audio files under 5 minutes. Non-commercial use only, audio may be collected. No API key required.

```APIDOC
## POST /process_audio

### Description
Processes an audio file using the VoiceFixer API.

### Method
POST

### Endpoint
https://voicefixer-voicefixer-api.hf.space/process_audio

### Parameters
#### Request Body
- **file** (multipart/form-data) - Required - The audio file to process. Can be specified using `@` notation (e.g., `file=@test.mp3`).

### Request Example
```bash
curl -X POST -H "Content-Type: multipart/form-data" -F "file=@test.mp3" https://voicefixer-voicefixer-api.hf.space/process_audio > processed_audio.wav
```

### Response
#### Success Response (200)
- **audio** (file) - The processed audio file in WAV format.
```

--------------------------------

### Process a single audio file using CLI

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Use the VoiceFixer command-line interface to process a single audio file. The output will be saved to 'outfile.wav' by default if no output path is specified.

```bash
voicefixer --infile test/utterance/original/original.wav
```

--------------------------------

### Use VoiceFixer API via cURL

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

This command demonstrates how to use the VoiceFixer API for audio processing via cURL. It uploads an MP3 file and saves the processed audio to a WAV file. Ensure the API endpoint is correct.

```bash
curl -X POST -H "Content-Type: multipart/form-data" -F "file=@test.mp3" https://voicefixer-voicefixer-api.hf.space/process_audio > processed_audio.wav
```

--------------------------------

### Vocoder Initialization and Forward Pass

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Initializes the neural vocoder and synthesizes speech waveforms from mel spectrograms.

```APIDOC
## `Vocoder.__init__(sample_rate)` and `Vocoder.forward(mel, cuda=False)`

### Description
Initialize and run the standalone neural vocoder on a mel spectrogram tensor.

`Vocoder` is a standalone `torch.nn.Module` that synthesizes speech waveforms from 128-band mel spectrograms. It is loaded separately from `VoiceFixer` and can be used independently to convert any mel spectrogram into audio. The vocoder is pre-trained and all parameters are frozen (no gradients). It expects non-normalized mel spectrograms of shape `[batch, 1, time_steps, 128]`.

### Method
`__init__`, `forward`

### Parameters
#### `__init__` Parameters
- **sample_rate** (int) - The sample rate for the vocoder.

#### `forward` Parameters
- **mel** (torch.Tensor) - The mel spectrogram tensor of shape `[batch, 1, time_steps, 128]`.
- **cuda** (bool, optional) - Whether to use GPU acceleration. Defaults to False.

### Request Example
```python
import torch
import librosa
import numpy as np
from voicefixer import Vocoder

# Initialize vocoder at 44.1 kHz
vocoder = Vocoder(sample_rate=44100)

# Create a dummy mel spectrogram (batch=1, channels=1, time=100, mel_bins=128)
mel = torch.rand(1, 1, 100, 128)  # replace with real mel spectrogram

# Synthesize waveform
with torch.no_grad():
    waveform = vocoder(mel, cuda=False)   # returns [1, 1, num_samples]

print(f"Output waveform shape: {waveform.shape}")
# Output: torch.Size([1, 1, 45056])

# GPU inference on CUDA
if torch.cuda.is_available():
    waveform_gpu = vocoder(mel, cuda=True)
```

### Response
- **waveform** (torch.Tensor) - The synthesized waveform tensor of shape `[batch, 1, num_samples]`.
```

--------------------------------

### Train VoiceFixer2 with Different Modes

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Iterates through different modes (0, 1, 2) to restore audio. Specify input/output paths and GPU acceleration. Note that mode 2 might not always work on degraded speech.

```python
for mode in [0,1,2]:
    print("Testing mode",mode)
    voicefixer.restore(
        input=os.path.join(git_root,"test/utterance/original/original.flac"), # low quality .wav/.flac file
        output=os.path.join(git_root,"test/utterance/output/output_mode_"+str(mode)+".flac"), # save file path
        cuda=False, # GPU acceleration
        mode=mode
    )
    if (mode != 2):
        check("output_mode_" + str(mode) + ".flac")
    print("Pass")
```

--------------------------------

### Restore WAV file via curl

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Use this curl command to restore a WAV file. The restored audio will be saved as a WAV file.

```bash
curl -X POST \
  -H "Content-Type: multipart/form-data" \
  -F "file=@degraded_speech.wav" \
  https://voicefixer-voicefixer-api.hf.space/process_audio \
  > restored_speech.wav
```

--------------------------------

### VoiceFixer CLI

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Command-line interface for restoring speech using VoiceFixer.

```APIDOC
## Command-Line Interface (CLI)

### Description
Restore speech from the terminal using the `voicefixer` command.

The CLI is installed automatically with the package as a console script entry point. It supports processing a single file or an entire folder, selecting the restoration mode, enabling GPU, and suppressing output. By default, if no output path is specified, output is saved to `outfile.wav`.

### Usage Examples
```bash
# Restore a single file (default mode 0, output to outfile.wav)
voicefixer --infile degraded_speech.wav

# Specify output path and mode
voicefixer --infile degraded_speech.wav --outfile restored.wav --mode 1

# Run all three modes (output files are suffixed with -mode0, -mode1, -mode2)
voicefixer --infile degraded_speech.wav --outfile restored.wav --mode all

# Process all files in a folder
voicefixer --infolder /path/to/input/ --outfolder /path/to/output/

# Disable GPU acceleration explicitly
voicefixer --infile degraded_speech.wav --outfile restored.wav --disable-cuda

# Suppress all console output
voicefixer --infile degraded_speech.wav --silent

# Show help
voicefixer -h
```
```

--------------------------------

### Python API - Processing Modes

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Explanation of the different processing modes available in the VoiceFixer Python API.

```APIDOC
### VoiceFixer Modes

- **Mode 0**: Original Model (suggested by default).
- **Mode 1**: Add preprocessing module (removes higher frequencies).
```

--------------------------------

### Restore MP3 file via curl

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Use this curl command to restore an MP3 file. The processed audio will be returned as a WAV file.

```bash
curl -X POST \
  -H "Content-Type: multipart/form-data" \
  -F "file=@test.mp3" \
  https://voicefixer-voicefixer-api.hf.space/process_audio \
  > processed_audio.wav
```

--------------------------------

### VoiceFixer CLI: Restore Speech from Terminal

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Command-line interface for VoiceFixer to restore speech from audio files or folders. Supports specifying input/output paths, restoration modes, and GPU acceleration.

```bash
# Restore a single file (default mode 0, output to outfile.wav)
voicefixer --infile degraded_speech.wav

# Specify output path and mode
voicefixer --infile degraded_speech.wav --outfile restored.wav --mode 1

# Run all three modes (output files are suffixed with -mode0, -mode1, -mode2)
voicefixer --infile degraded_speech.wav --outfile restored.wav --mode all

# Process all files in a folder
voicefixer --infolder /path/to/input/ --outfolder /path/to/output/

# Disable GPU acceleration explicitly
voicefixer --infile degraded_speech.wav --outfile restored.wav --disable-cuda

# Suppress all console output
voicefixer --infile degraded_speech.wav --silent

# Show help
voicefixer -h
```

--------------------------------

### Restore Speech Audio File

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Restore a degraded speech audio file and save the result to disk. Supports various audio formats and offers different restoration modes. GPU acceleration can be enabled.

```python
import os
from voicefixer import VoiceFixer

voicefixer = VoiceFixer()

# Mode 0: Original model — best for general use
voicefixer.restore(
    input="degraded_speech.wav",   # input: low-quality .wav or .flac
    output="restored_mode0.wav",   # output: restored .wav file
    cuda=False,                    # set True for GPU (CUDA or MPS)
    mode=0
)

# Mode 1: Preprocessing removes high-frequency noise before restoration
voicefixer.restore(
    input="noisy_recording.flac",
    output="restored_mode1.wav",
    cuda=True,   # GPU acceleration (NVIDIA CUDA or Apple M1/M2 via MPS)
    mode=1
)

# Mode 2: Train mode — may help with severely clipped or degraded speech
voicefixer.restore(
    input="old_recording.wav",
    output="restored_mode2.wav",
    cuda=False,
    mode=2
)

# Process all three modes at once (output files get -mode0, -mode1, -mode2 suffixes)
for mode in [0, 1, 2]:
    voicefixer.restore(
        input="degraded_speech.wav",
        output=f"restored_mode{mode}.wav",
        cuda=False,
        mode=mode
    )
```

--------------------------------

### VoiceFixer.restore

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Restores a degraded speech audio file and writes the result to disk. This is the primary high-level API for file-based speech restoration.

```APIDOC
## VoiceFixer.restore(input, output, cuda=False, mode=0, your_vocoder_func=None, tqdm=tqdm)

### Description
Restore a degraded speech audio file and write the result to disk

This is the primary high-level API for file-based speech restoration. It loads the input audio at 44.1 kHz using librosa, processes it through the neural restoration pipeline in 30-second segments, and saves the restored audio to the output path. It supports `.wav`, `.flac`, and (with FFmpeg) `.mp3` and other formats. The `cuda` flag enables GPU acceleration on NVIDIA GPUs and Apple Silicon (MPS). A custom vocoder function can be passed via `your_vocoder_func` to replace the built-in neural vocoder.

### Parameters
#### Path Parameters
- **input** (string) - Required - Path to the input audio file (.wav, .flac, .mp3, etc.).
- **output** (string) - Required - Path to save the restored audio file.
#### Query Parameters
- **cuda** (boolean) - Optional - Set to True for GPU acceleration (NVIDIA CUDA or Apple M1/M2 via MPS). Defaults to False.
- **mode** (integer) - Optional - Restoration mode. 0: original model (recommended), 1: high-frequency preprocessing, 2: training mode. Defaults to 0.
- **your_vocoder_func** (function) - Optional - A custom vocoder function to replace the built-in neural vocoder.
- **tqdm** (object) - Optional - TQDM progress bar object.

### Request Example
```python
import os
from voicefixer import VoiceFixer

voicefixer = VoiceFixer()

# Mode 0: Original model — best for general use
voicefixer.restore(
    input="degraded_speech.wav",   # input: low-quality .wav or .flac
    output="restored_mode0.wav",   # output: restored .wav file
    cuda=False,                    # set True for GPU (CUDA or MPS)
    mode=0
)

# Mode 1: Preprocessing removes high-frequency noise before restoration
voicefixer.restore(
    input="noisy_recording.flac",
    output="restored_mode1.wav",
    cuda=True,   # GPU acceleration (NVIDIA CUDA or Apple M1/M2 via MPS)
    mode=1
)

# Mode 2: Train mode — may help with severely clipped or degraded speech
voicefixer.restore(
    input="old_recording.wav",
    output="restored_mode2.wav",
    cuda=False,
    mode=2
)

# Process all three modes at once (output files get -mode0, -mode1, -mode2 suffixes)
for mode in [0, 1, 2]:
    voicefixer.restore(
        input="degraded_speech.wav",
        output=f"restored_mode{mode}.wav",
        cuda=False,
        mode=mode
    )
```
```

--------------------------------

### VoiceFixer.restore_inmem

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Restores speech from an in-memory numpy waveform array, returning a numpy array without touching the filesystem. This is a lower-level API useful for integrating into audio processing pipelines.

```APIDOC
## VoiceFixer.restore_inmem(wav_10k, cuda=False, mode=0, your_vocoder_func=None, tqdm=tqdm)

### Description
Restore speech from an in-memory numpy waveform array, returning a numpy array

This lower-level API accepts a pre-loaded numpy waveform (resampled to 44.1 kHz) and returns a restored numpy array without touching the filesystem. It is useful when integrating VoiceFixer into audio processing pipelines (e.g., with pydub, soundfile, or torchaudio). Audio is processed in 30-second chunks with automatic frame alignment and energy normalization.

### Parameters
#### Path Parameters
- **wav_10k** (numpy.ndarray) - Required - Pre-loaded numpy waveform array (resampled to 44.1 kHz).
#### Query Parameters
- **cuda** (boolean) - Optional - Set to True for GPU acceleration (NVIDIA CUDA or Apple M1/M2 via MPS). Defaults to False.
- **mode** (integer) - Optional - Restoration mode. 0: original model (recommended), 1: high-frequency preprocessing, 2: training mode. Defaults to 0.
- **your_vocoder_func** (function) - Optional - A custom vocoder function to replace the built-in neural vocoder.
- **tqdm** (object) - Optional - TQDM progress bar object.

### Request Example
```python
import librosa
import soundfile as sf
from voicefixer import VoiceFixer

voicefixer = VoiceFixer()

# Example usage (assuming you have a degraded audio file loaded as a numpy array)
# degraded_audio_path = "degraded_speech.wav"
# wav_10k, sr = librosa.load(degraded_audio_path, sr=44100)

# restored_wav = voicefixer.restore_inmem(wav_10k, cuda=False, mode=0)

# To save the restored audio to a file:
# sf.write("restored_inmem.wav", restored_wav, sr)
```
```

--------------------------------

### Vocoder Oracle Test: Reconstruct Audio from Mel Spectrogram

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Use the `Vocoder.oracle` method to test vocoder quality by reconstructing audio from a clean reference file's mel spectrogram. This establishes an upper bound on vocoder fidelity.

```python
from voicefixer import Vocoder

vocoder = Vocoder(sample_rate=44100)

# Reconstruct from oracle mel spectrogram (reads input, synthesizes, saves output)
vocoder.oracle(
    fpath="clean_reference.flac",
    out_path="oracle_reconstruction.flac",
    cuda=False
)

# With GPU acceleration
vocoder.oracle(
    fpath="clean_reference.flac",
    out_path="oracle_reconstruction_gpu.flac",
    cuda=True
)
```

--------------------------------

### Vocoder Oracle Test

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Reconstructs audio from a clean waveform via ground-truth mel spectrogram for vocoder quality testing.

```APIDOC
## `Vocoder.oracle(fpath, out_path, cuda=False)`

### Description
Reconstruct audio from a clean waveform via ground-truth mel spectrogram (vocoder oracle test).

This method reads a clean reference audio file, computes its mel spectrogram using librosa, feeds it through the neural vocoder, and saves the result. It is primarily used for testing vocoder quality without the upstream restoration model, establishing an upper-bound on vocoder fidelity.

### Method
`oracle`

### Parameters
- **fpath** (str) - Path to the input clean audio file.
- **out_path** (str) - Path to save the reconstructed audio file.
- **cuda** (bool, optional) - Whether to use GPU acceleration. Defaults to False.

### Request Example
```python
from voicefixer import Vocoder

vocoder = Vocoder(sample_rate=44100)

# Reconstruct from oracle mel spectrogram (reads input, synthesizes, saves output)
vocoder.oracle(
    fpath="clean_reference.flac",
    out_path="oracle_reconstruction.flac",
    cuda=False
)

# With GPU acceleration
vocoder.oracle(
    fpath="clean_reference.flac",
    out_path="oracle_reconstruction_gpu.flac",
    cuda=True
)
```
```

--------------------------------

### Restore Speech In-Memory

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Restore speech from an in-memory numpy waveform array and return a numpy array. This is useful for integrating into audio processing pipelines.

```python
import librosa
import soundfile as sf
from voicefixer import VoiceFixer

voicefixer = VoiceFixer()


```

--------------------------------

### Change VoiceFixer mode via CLI

Source: https://github.com/render-ai-team/voicefixer2/blob/main/README.md

Specify a processing mode (e.g., mode 1 for preprocessing) when using the VoiceFixer CLI. The default mode is 0. This allows for different audio restoration strategies.

```bash
voicefixer --infile /path/to/input.wav --outfile /path/to/output.wav --mode 1
```

--------------------------------

### Remove Higher Frequency Content with VoiceFixer

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Use VoiceFixer's `remove_higher_frequency` utility to filter out high-frequency noise from an audio waveform. This can be used as a standalone pre-filter.

```python
import librosa
import soundfile as sf
from voicefixer import VoiceFixer

voicefixer = VoiceFixer()

# Load raw waveform
wav, _ = librosa.load("high_freq_noise.wav", sr=44100)

# Remove frequencies above 95% energy threshold
filtered_wav = voicefixer.remove_higher_frequency(wav, ratio=0.95)

# Save filtered audio
sf.write("filtered_output.wav", filtered_wav, samplerate=44100)
print(f"Original length: {len(wav)}, Filtered length: {len(filtered_wav)}")
```

--------------------------------

### VoiceFixer.remove_higher_frequency

Source: https://context7.com/render-ai-team/voicefixer2/llms.txt

Removes high-frequency content from a waveform via STFT-based spectral truncation. This can be used as a standalone noise reduction pre-filter.

```APIDOC
## `VoiceFixer.remove_higher_frequency(wav, ratio=0.95)`

### Description
Remove high-frequency content from a waveform via STFT-based spectral truncation

This preprocessing utility zeroes out frequency bins beyond the energy threshold defined by `ratio`. It is applied automatically during Mode 1 restoration on each 30-second segment. It can also be used standalone as a noise reduction pre-filter before passing audio to other tools.

### Method
`remove_higher_frequency`

### Parameters
- **wav** (np.ndarray) - The input waveform.
- **ratio** (float, optional) - The energy threshold ratio. Defaults to 0.95.

### Request Example
```python
import librosa
import soundfile as sf
from voicefixer import VoiceFixer

voicefixer = VoiceFixer()

# Load raw waveform
wav, _ = librosa.load("high_freq_noise.wav", sr=44100)

# Remove frequencies above 95% energy threshold
filtered_wav = voicefixer.remove_higher_frequency(wav, ratio=0.95)

# Save filtered audio
sf.write("filtered_output.wav", filtered_wav, samplerate=44100)
print(f"Original length: {len(wav)}, Filtered length: {len(filtered_wav)}")
```

### Response
- **filtered_wav** (np.ndarray) - The waveform with high-frequency content removed.
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.