Spaces:
Running
Running
| import gradio as gr | |
| import torch | |
| from torchaudio.backend.common import AudioMetaData | |
| from df.enhance import enhance, load_audio, save_audio | |
| from df.io import resample | |
| from libdf import DF | |
| from df.model import ModelParams | |
| from df import config | |
| import moviepy.editor as mp | |
| import numpy as np | |
| try: | |
| config.load('config.ini') | |
| except Exception as e: | |
| print(e) | |
| DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu') | |
| p = ModelParams() | |
| df = DF( | |
| sr=p.sr, | |
| fft_size=p.fft_size, | |
| hop_size=p.hop_size, | |
| nb_bands=p.nb_erb, | |
| min_nb_erb_freqs=p.min_nb_freqs, | |
| ) | |
| print("Device - ", DEVICE) | |
| model = torch.load(("model.pth"), map_location=torch.device('cpu')) | |
| model.to(DEVICE) | |
| model.eval() | |
| def identity(video_path): | |
| print(video_path) | |
| # audio = mp.AudioFileClip(x) | |
| # wav_file = x | |
| # audio.write_audiofile(wav_file) | |
| video = mp.VideoFileClip(video_path) | |
| audio = video.audio | |
| wav_file = "tmp.wav" | |
| audio.write_audiofile(wav_file) | |
| print("Wav stored.") | |
| meta = AudioMetaData(-1, -1, -1, -1, "") | |
| sr = config("sr", 48000, int, section="df") | |
| sample, meta = load_audio(wav_file, sr) | |
| len_audio = (meta.num_frames/meta.sample_rate)/60 | |
| max_min = 1 | |
| if len_audio % max_min < 0.1: | |
| num_chunks = len_audio // max_min | |
| else: | |
| num_chunks = len_audio // max_min + 1 | |
| if len_audio and not num_chunks: | |
| num_chunks += 1 | |
| print(f"Total length of audio = {len_audio} chunks = {num_chunks}") | |
| estimate = [] | |
| split_tensors = torch.tensor_split(sample, int(num_chunks), dim = 1) | |
| for i in range(len(split_tensors)): | |
| enhanced = enhance(model, df, split_tensors[i]) | |
| enhanced = enhance(model, df, enhanced) | |
| lim = torch.linspace(0.0, 1.0, int(sr * 0.15)).unsqueeze(0) | |
| lim = torch.cat((lim, torch.ones(1, enhanced.shape[1] - lim.shape[1])), dim=1) | |
| enhanced = enhanced * lim | |
| enhanced = resample(enhanced, sr, meta.sample_rate) | |
| estimate.append(enhanced) | |
| estimate = tuple(estimate) | |
| enhanced = torch.cat(estimate, dim = -1) | |
| sr = meta.sample_rate | |
| save_audio("enhanced_aud.wav", enhanced, sr) | |
| audio = mp.AudioFileClip('enhanced_aud.wav') | |
| video = mp.VideoFileClip(video_path) | |
| final_video = video.set_audio(audio) | |
| final_video.write_videofile("output_video.mp4", | |
| codec='libx264', | |
| audio_codec='aac', | |
| temp_audiofile='temp-audio.m4a', | |
| remove_temp=True | |
| ) | |
| return "output_video.mp4" | |
| demo = gr.Interface( | |
| fn=identity, | |
| title="Audio Denoiser using DeepFilterNet V3", | |
| description="Implemented audio denoising using DeepFilterNet V3, enabled processing of larger files even on cpu, by splitting up the audio file into chunks of 1 minute each.\n\nThe processing will be very slow since it's the free version of HuggingFace, 2 second audio can take about 5 minutes.", | |
| inputs=gr.Video(label="Input Video", sources="upload"), | |
| outputs=gr.Video(label="Output Video"), | |
| ) | |
| demo.launch() |