import gradio as gr # import os, subprocess, torchaudio # import torch from PIL import Image from gtts import gTTS import tempfile from pydub import AudioSegment from pydub.generators import Sine # from fairseq.checkpoint_utils import load_model_ensemble_and_task_from_hf_hub # from fairseq.models.text_to_speech.hub_interface import TTSHubInterface import soundfile import dlib import cv2 import imageio import os import gradio as gr import os, subprocess, torchaudio from PIL import Image import ffmpeg block = gr.Blocks() def one_shot_talking(image_in,audio_in): # image = Image.open(image_in) # image = crop_src_image(image) # image.save("image_pre.png") # #Pre-processing of image # # crop_src_image(image_in) # exit() #Improve quality of input image # os.system(f"python /content/GFPGAN/inference_gfpgan.py --upscale 2 -i /content/image_pre.png -o /content/results --bg_upsampler realesrgan") # image_in_one_shot='/content/results/restored_imgs/image_pre.png' waveform, sample_rate = torchaudio.load(audio_in) torchaudio.save("/content/audio.wav", waveform, sample_rate, encoding="PCM_S", bits_per_sample=16) image = Image.open(image_in) image = pad_image(image) image.save("/content/image_pre.png") return "/content/audio.wav" pocketsphinx_run = subprocess.run(['pocketsphinx', '-phone_align', 'yes', 'single', '/content/audio.wav'], check=True, capture_output=True) jq_run = subprocess.run(['jq', '[.w[]|{word: (.t | ascii_upcase | sub(""; "sil") | sub(""; "sil") | sub("\\\(2\\\)"; "") | sub("\\\(3\\\)"; "") | sub("\\\(4\\\)"; "") | sub("\\\[SPEECH\\\]"; "SIL") | sub("\\\[NOISE\\\]"; "SIL")), phones: [.w[]|{ph: .t | sub("\\\+SPN\\\+"; "SIL") | sub("\\\+NSN\\\+"; "SIL"), bg: (.b*100)|floor, ed: (.b*100+.d*100)|floor}]}]'], input=pocketsphinx_run.stdout, capture_output=True) with open("test.json", "w") as f: f.write(jq_run.stdout.decode('utf-8').strip()) os.system(f"cd /content/one-shot-talking-face && python3 -B test_script.py --img_path /content/image_pre.png --audio_path /content/audio.wav --phoneme_path /content/test.json --save_dir /content/train") exit() #Video Quality Improvement #1. Extract the frames from the video file using PyVideoFramesExtractor os.system(f"python /content/PyVideoFramesExtractor/extract.py --video=/content/train/image_pre_audio.mp4") #2. Improve image quality using GFPGAN on each frames os.system(f"python /content/GFPGAN/inference_gfpgan.py --upscale 2 -i /content/extracted_frames/ -o /content/video_results --bg_upsampler realesrgan") #3. Merge all the frames to a one video using imageio path = '/content/video_results/restored_imgs' image_folder = os.fsencode(path) print(image_folder) filenames = [] for file in os.listdir(image_folder): filename = os.fsdecode(file) if filename.endswith( ('.jpg', '.png', '.gif') ): filenames.append(filename) filenames.sort() # this iteration technique has no built in order, so sort the frames images = list(map(lambda filename: imageio.imread("/content/video_results/restored_imgs/"+filename), filenames)) imageio.mimsave('/content/video_output.mp4', images, fps=25.0) # modify the frame duration as needed input_video = ffmpeg.input('/content/video_output.mp4') input_audio = ffmpeg.input('/content/audio.wav') ffmpeg.concat(input_video, input_audio, v=1, a=1).output('final_output.mp4').run() return "final_output.mp4" def one_shot(image,input_text,gender): if gender == 'Female' or gender == 'female': tts = gTTS(input_text) with tempfile.NamedTemporaryFile(suffix='.mp3', delete=False) as f: tts.write_to_fp(f) f.seek(0) sound = AudioSegment.from_file(f.name, format="mp3") sound.export("/content/audio.wav", format="wav") waveform, sample_rate = torchaudio.load(audio_in) torchaudio.save("/content/audio.wav", waveform, sample_rate, encoding="PCM_S", bits_per_sample=16) image = Image.open(image_in) image = pad_image(image) image.save("/content/image_pre.png") pocketsphinx_run = subprocess.run(['pocketsphinx', '-phone_align', 'yes', 'single', '/content/audio.wav'], check=True, capture_output=True) jq_run = subprocess.run(['jq', '[.w[]|{word: (.t | ascii_upcase | sub(""; "sil") | sub(""; "sil") | sub("\\\(2\\\)"; "") | sub("\\\(3\\\)"; "") | sub("\\\(4\\\)"; "") | sub("\\\[SPEECH\\\]"; "SIL") | sub("\\\[NOISE\\\]"; "SIL")), phones: [.w[]|{ph: .t | sub("\\\+SPN\\\+"; "SIL") | sub("\\\+NSN\\\+"; "SIL"), bg: (.b*100)|floor, ed: (.b*100+.d*100)|floor}]}]'], input=pocketsphinx_run.stdout, capture_output=True) with open("test.json", "w") as f: f.write(jq_run.stdout.decode('utf-8').strip()) os.system(f"cd /content/one-shot-talking-face && python3 -B test_script.py --img_path /content/image_pre.png --audio_path /content/audio.wav --phoneme_path /content/test.json --save_dir /content/train") #1. Extract the frames from the video file using PyVideoFramesExtractor os.system(f"python /content/PyVideoFramesExtractor/extract.py --video=/content/train/image_pre_audio.mp4") #2. Improve image quality using GFPGAN on each frames os.system(f"python /content/GFPGAN/inference_gfpgan.py --upscale 2 -i /content/extracted_frames/ -o /content/video_results --bg_upsampler realesrgan") #3. Merge all the frames to a one video using imageio path = '/content/video_results/restored_imgs' image_folder = os.fsencode(path) print(image_folder) filenames = [] return "/content/audio.wav" elif gender == 'Male' or gender == 'male': print(gender) models, cfg, task = load_model_ensemble_and_task_from_hf_hub( "Voicemod/fastspeech2-en-male1", arg_overrides={"vocoder": "hifigan", "fp16": False} ) model = models[0].cuda() TTSHubInterface.update_cfg_with_data_cfg(cfg, task.data_cfg) generator = task.build_generator([model], cfg) # next(model.parameters()).device sample = TTSHubInterface.get_model_input(task, input_text) sample["net_input"]["src_tokens"] = sample["net_input"]["src_tokens"].cuda() sample["net_input"]["src_lengths"] = sample["net_input"]["src_lengths"].cuda() sample["speaker"] = sample["speaker"].cuda() wav, rate = TTSHubInterface.get_prediction(task, model, generator, sample) # soundfile.write("/content/audio_before.wav", wav, rate) soundfile.write("/content/audio_before.wav", wav.cpu().clone().numpy(), rate) cmd='ffmpeg -i /content/audio_before.wav -filter:a "atempo=0.7" -vn /content/audio.wav' os.system(cmd) one_shot_talking(image,'audio.wav') def generate_ocr(method,image,gender): return "Hello" def run(): with block: with gr.Group(): with gr.Box(): with gr.Row().style(equal_height=True): image_in = gr.Image(show_label=False, type="filepath") # audio_in = gr.Audio(show_label=False, type='filepath') input_text=gr.Textbox(lines=3, value="Hello How are you?", label="Input Text") gender = gr.Radio(["Female","Male"],value="Female",label="Gender") video_out = gr.Audio(label="output") # video_out = gr.Video(show_label=False) with gr.Row().style(equal_height=True): btn = gr.Button("Generate") btn.click(one_shot, inputs=[image_in, input_text,gender], outputs=[video_out]) # block.queue() block.launch(server_name="0.0.0.0", server_port=7860) if __name__ == "__main__": run()