diff --git a/codes/scripts/audio/gen/use_diffuse_tts.py b/codes/scripts/audio/gen/use_diffuse_tts.py index 0984365f..08a0fbec 100644 --- a/codes/scripts/audio/gen/use_diffuse_tts.py +++ b/codes/scripts/audio/gen/use_diffuse_tts.py @@ -112,17 +112,17 @@ if __name__ == '__main__': ] parser = argparse.ArgumentParser() - parser.add_argument('-text', type=str, help='Text to speak.', default='This is the real secret of life: to be completely engaged in what you are doing in the here and now. To realize that instead of work, it is play.') - parser.add_argument('-opt_code_gen', type=str, help='Path to options YAML file used to train the code_gen model', default='D:\\dlas\\options\\train_encoder_build_ctc_alignments.yml') + parser.add_argument('-text', type=str, help='Text to speak.', default='instead of molten iron, jupiter and brown dwarfs have hydrogen, which is under so much pressure that it develops metallic properties.') + parser.add_argument('-opt_code_gen', type=str, help='Path to options YAML file used to train the code_gen model', default='D:\\dlas\\experiments\\train_encoder_build_ctc_alignments_medium\\train_encoder_build_ctc_alignments.yml') parser.add_argument('-code_gen_model_name', type=str, help='Name of the code_gen model in opt.', default='generator') - parser.add_argument('-code_gen_model_path', type=str, help='Path to saved code_gen model weights', default='D:\\dlas\\experiments\\train_encoder_build_ctc_alignments\\models\\31000_generator_ema.pth') + parser.add_argument('-code_gen_model_path', type=str, help='Path to saved code_gen model weights', default='D:\\dlas\\experiments\\train_encoder_build_ctc_alignments_medium\\models\\32000_generator_ema.pth') parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', default='X:\\dlas\\experiments\\train_diffusion_tts5_medium.yml') parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator') parser.add_argument('-diffusion_model_path', type=str, help='Path to saved model weights', default='X:\\dlas\\experiments\\train_diffusion_tts5_medium\\models\\73000_generator_ema.pth') parser.add_argument('-sr_opt', type=str, help='Path to options YAML file used to train the SR diffusion model', default='X:\\dlas\\experiments\\train_diffusion_tts6_upsample.yml') parser.add_argument('-sr_diffusion_model_name', type=str, help='Name of the SR diffusion model in opt.', default='generator') - parser.add_argument('-sr_diffusion_model_path', type=str, help='Path to saved model weights for the SR diffuser', default='X:\\dlas\\experiments\\train_diffusion_tts6_upsample\\models\\31000_generator_ema.pth') - parser.add_argument('-cond', type=str, help='Type of conditioning voice', default='simmons') + parser.add_argument('-sr_diffusion_model_path', type=str, help='Path to saved model weights for the SR diffuser', default='X:\\dlas\\experiments\\train_diffusion_tts6_upsample_continued\\models\\45000_generator_ema.pth') + parser.add_argument('-cond', type=str, help='Type of conditioning voice', default='the_doctor') parser.add_argument('-diffusion_steps', type=int, help='Number of diffusion steps to perform to create the generate. Lower steps reduces quality, but >40 is generally pretty good.', default=100) parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='../results/use_diffuse_tts') parser.add_argument('-device', type=str, help='Device to run on', default='cuda') diff --git a/codes/scripts/audio/gen/use_diffuse_voice_translation.py b/codes/scripts/audio/gen/use_diffuse_voice_translation.py index a333d6e9..e5d8c209 100644 --- a/codes/scripts/audio/gen/use_diffuse_voice_translation.py +++ b/codes/scripts/audio/gen/use_diffuse_voice_translation.py @@ -5,6 +5,7 @@ import torch import torchaudio from data.audio.unsupervised_audio_dataset import load_audio +from data.util import find_files_of_type, is_audio_file from scripts.audio.gen.speech_synthesis_utils import do_spectrogram_diffusion, \ load_discrete_vocoder_diffuser, wav_to_mel, convert_mel_to_codes from utils.audio import plot_spectrogram @@ -47,14 +48,14 @@ if __name__ == '__main__': } parser = argparse.ArgumentParser() - parser.add_argument('-src_clip', type=str, help='Path to the audio file to translate', default='D:\\tortoise-tts\\voices\\dotrice\\1.wav') + parser.add_argument('-src_clip', type=str, help='Path to the audio files to translate', default='D:\\tortoise-tts\\voices') parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', default='X:\\dlas\\experiments\\train_diffusion_tts5_medium.yml') parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator') parser.add_argument('-diffusion_model_path', type=str, help='Path to saved model weights', default='X:\\dlas\\experiments\\train_diffusion_tts5_medium\\models\\73000_generator_ema.pth') parser.add_argument('-sr_opt', type=str, help='Path to options YAML file used to train the SR diffusion model', default='X:\\dlas\\experiments\\train_diffusion_tts6_upsample.yml') parser.add_argument('-sr_diffusion_model_name', type=str, help='Name of the SR diffusion model in opt.', default='generator') - parser.add_argument('-sr_diffusion_model_path', type=str, help='Path to saved model weights for the SR diffuser', default='X:\\dlas\\experiments\\train_diffusion_tts6_upsample_continued\\models\\41000_generator_ema.pth') - parser.add_argument('-voice', type=str, help='Type of conditioning voice', default='puppy') + parser.add_argument('-sr_diffusion_model_path', type=str, help='Path to saved model weights for the SR diffuser', default='X:\\dlas\\experiments\\train_diffusion_tts6_upsample_continued\\models\\45000_generator_ema.pth') + parser.add_argument('-voice', type=str, help='Type of conditioning voice', default='simmons') parser.add_argument('-diffusion_steps', type=int, help='Number of diffusion steps to perform to create the generate. Lower steps reduces quality, but >40 is generally pretty good.', default=100) parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='../results/use_diffuse_voice_translation') parser.add_argument('-device', type=str, help='Device to run on', default='cuda') @@ -79,25 +80,30 @@ if __name__ == '__main__': torchaudio.save(os.path.join(args.output_path, 'cond_sr.wav'), sr_cond.cpu(), sr_sample_rate) with torch.no_grad(): - print("Extracting CTC codes from source clip..") - aligned_codes, src_clip = get_ctc_codes_for(args.src_clip) - torchaudio.save(os.path.join(args.output_path, f'source_clip.wav'), src_clip.unsqueeze(0).cpu(), 16000) + if os.path.isdir(args.src_clip): + files = find_files_of_type('img', args.src_clip, qualifier=is_audio_file)[0] + else: + files = [args.src_clip] + for e, file in enumerate(files): + print("Extracting CTC codes from source clip..") + aligned_codes, src_clip = get_ctc_codes_for(file) + torchaudio.save(os.path.join(args.output_path, f'{e}_source_clip.wav'), src_clip.unsqueeze(0).cpu(), 16000) - print("Performing initial diffusion..") - output_shape = (1, 1, ceil_multiple(aligned_codes.shape[-1]*aligned_codes_compression_factor, 2048)) - diffusion = diffusion.cuda() - output_base = diffuser.p_sample_loop(diffusion, output_shape, noise=torch.zeros(output_shape, device=args.device), - model_kwargs={'tokens': aligned_codes, - 'conditioning_input': cond.unsqueeze(0)}) - diffusion = diffusion.cpu() - torchaudio.save(os.path.join(args.output_path, f'output_mean_base.wav'), output_base.cpu().squeeze(0), base_sample_rate) + print("Performing initial diffusion..") + output_shape = (1, 1, ceil_multiple(aligned_codes.shape[-1]*aligned_codes_compression_factor, 2048)) + diffusion = diffusion.cuda() + output_base = diffuser.p_sample_loop(diffusion, output_shape, noise=torch.zeros(output_shape, device=args.device), + model_kwargs={'tokens': aligned_codes, + 'conditioning_input': cond.unsqueeze(0)}) + diffusion = diffusion.cpu() + torchaudio.save(os.path.join(args.output_path, f'{e}_output_mean_base.wav'), output_base.cpu().squeeze(0), base_sample_rate) - print("Performing SR diffusion..") - output_shape = (1, 1, output_base.shape[-1] * (sr_sample_rate // base_sample_rate)) - sr_diffusion = sr_diffusion.cuda() - output = sr_diffuser.p_sample_loop(sr_diffusion, output_shape, noise=torch.zeros(output_shape, device=args.device), - model_kwargs={'tokens': aligned_codes, - 'conditioning_input': sr_cond.unsqueeze(0), - 'lr_input': output_base}) - sr_diffusion = sr_diffusion.cpu() - torchaudio.save(os.path.join(args.output_path, f'output_mean_sr.wav'), output.cpu().squeeze(0), sr_sample_rate) + print("Performing SR diffusion..") + output_shape = (1, 1, output_base.shape[-1] * (sr_sample_rate // base_sample_rate)) + sr_diffusion = sr_diffusion.cuda() + output = sr_diffuser.p_sample_loop(sr_diffusion, output_shape, #noise=torch.zeros(output_shape, device=args.device), + model_kwargs={'tokens': aligned_codes, + 'conditioning_input': sr_cond.unsqueeze(0), + 'lr_input': output_base}) + sr_diffusion = sr_diffusion.cpu() + torchaudio.save(os.path.join(args.output_path, f'{e}_output_mean_sr.wav'), output.cpu().squeeze(0), sr_sample_rate)