MLX BigVGAN

An MLX-adapted implementation of BigVGAN.

Features

• BigVGAN Integration: Fully integrates the original BigVGAN model with MLX for enhanced compatibility and performance.
• Flexible Conversion: Includes tools to convert the original BigVGAN PyTorch weights to MLX format.
• Customizable Configurations: Supports various configurations for kernel sizes, dilation rates, and activation functions (e.g., snake, snakebeta).
• Pretrained Models: Easily load pretrained BigVGAN models from the Hugging Face Hub.

Installation

pip install mlx-bigvgan

Usage

1. Load Pretrained Model

from mlx_bigvgan import BigVGAN

model = BigVGAN.from_pretrained("wyrom/mlx-bigvgan_v2_24khz_100band_256x")
model.eval()
mx.eval(model.parameters())

2. Generate Audio

import numpy as np
import mlx.core as mx
from mlx_bigvgan import log_mel_spectrogram, load_audio
# Load audio file
audio = load_audio("path/to/audio.wav")
h = model.config
# Compute log-mel spectrogram
mel_spec = log_mel_spectrogram(audio,
    n_fft=h.n_fft,
    n_mels=h.num_mels,
    sample_rate=h.sampling_rate,
    hop_length=h.hop_size,
    fmin=h.fmin,
    fmax=h.fmax,
    padding=(h.n_fft - h.hop_size) // 2,
    mel_norm="slaney",
    mel_scale="slaney",
    power=1.0,
)
# reshape to [B(1), T, C_mels]
mel_spec = mx.expand_dims(mel_spec, 0)
# Generate waveform
waveform = model(mel_spec) # [B(1), T, 1]
# Reshape to [T, 1]
waveform_float = waveform.squeeze(0)

# Convert to int16
waveform_int16 = mx.clip(waveform_float * 32767, -32768, 32767).astype(mx.int16)

# save to wav
import soundfile as sf

sf.write("output.wav", waveform_int16, h.sampling_rate, "PCM_16")

3. Convert Original BigVGAN Weights to MLX Format

You can convert the original BigVGAN weights to MLX format using the provided script.

repo_id is the Hugging Face model ID of the original BigVGAN model you want to convert.

See nvidia/BigVGAN for move pretrained models.

python -m mlx_bigvgan.convert --repo_id nvidia/bigvgan_v2_xxx  --output_dir mlx_models

References

• BigVGAN
• MLX

License

This project is licensed under the MIT License. See the LICENSE file for details.