def join_audios(base_dir: Path, directory: Path, silence_files: float,
silence_directories: float, output_file: Path,
dirs_before_files: bool, overwrite: bool) -> bool:
logger = getLogger(__name__)
if not directory.is_dir():
logger.error("Directory was not found!")
return False
if output_file.exists() and not overwrite:
logger.error("File already exists!")
return False
all_files = get_files_sorted_recursive(directory, not dirs_before_files)
all_wav_files = tuple(
tuple(file for file in file_chunk if file.suffix.lower() == ".wav")
for file_chunk in all_files)
final_parts = []
global_sr = None
for chunk_nr, file_chunk in enumerate(all_wav_files, start=1):
if len(file_chunk) == 0:
continue
is_last_chunk = chunk_nr == len(all_wav_files)
for file_nr, file in enumerate(file_chunk, start=1):
is_last_file = file_nr == len(file_chunk)
wav, wav_sr = wav_to_float32(file)
if len(wav.shape) != 1:
logger.error(
f"File {file} is not mono. Only mono files are supported! Skipped."
)
continue
if global_sr is None:
global_sr = wav_sr
else:
if wav_sr != global_sr:
logger.error(
f"File {file} has another sampling rate than the first file. Skipped."
)
continue
final_parts.append(wav)
if not is_last_file and silence_files > 0:
pause_samples = np.zeros(
(get_sample_count(wav_sr, silence_files), ))
final_parts.append(pause_samples)
if not is_last_chunk and silence_directories > 0:
pause_samples = np.zeros((get_sample_count(wav_sr,
silence_directories), ))
final_parts.append(pause_samples)
logger.debug("Concatenating...")
result = np.concatenate(tuple(final_parts), axis=-1)
logger.debug("Saving...")
output_file.parent.mkdir(parents=True, exist_ok=True)
float_to_wav(result, output_file, sample_rate=global_sr)
logger.info(f"Done. Written output to: {output_file.absolute()}")
def save_results(output: InferenceEntryOutput, infer_dir: str,
denoised_audio_wav_paths: List[Dict[str, Any]]):
dest_dir = get_inferred_mel_dir(infer_dir, output.identifier)
os.makedirs(dest_dir, exist_ok=True)
imageio.imsave(os.path.join(dest_dir, "original.png"), output.mel_orig_img)
imageio.imsave(os.path.join(dest_dir, "inferred_denoised.png"),
output.mel_inferred_denoised_img)
imageio.imsave(os.path.join(dest_dir, "diff.png"),
output.mel_denoised_diff_img)
np.save(os.path.join(dest_dir, "original.mel.npy"), output.mel_orig)
np.save(os.path.join(dest_dir, "inferred_denoised.mel.npy"),
output.mel_inferred_denoised)
inferred_denoised_path = os.path.join(dest_dir, "inferred_denoised.wav")
float_to_wav(output.wav_inferred_denoised,
inferred_denoised_path,
sample_rate=output.inferred_sr)
float_to_wav(output.wav_inferred,
os.path.join(dest_dir, "inferred.wav"),
sample_rate=output.inferred_sr)
stack_images_vertically(list_im=[
os.path.join(dest_dir, "original.png"),
os.path.join(dest_dir, "inferred_denoised.png"),
os.path.join(dest_dir, "diff.png"),
],
out_path=os.path.join(dest_dir, "comparison.png"))
wav_info = get_wav_info_dict(
identifier=output.identifier,
path=inferred_denoised_path,
sr=output.inferred_sr,
)
denoised_audio_wav_paths.append(wav_info)
def save_results(entry: PreparedData, output: ValidationEntryOutput, val_dir: str, iteration: int):
dest_dir = get_val_entry_dir(val_dir, entry, iteration)
imageio.imsave(os.path.join(dest_dir, "original.png"), output.mel_orig_img)
imageio.imsave(os.path.join(dest_dir, "inferred_denoised.png"), output.mel_inferred_denoised_img)
imageio.imsave(os.path.join(dest_dir, "diff.png"), output.mel_denoised_diff_img)
np.save(os.path.join(dest_dir, "original.mel.npy"), output.mel_orig)
np.save(os.path.join(dest_dir, "inferred_denoised.mel.npy"), output.mel_inferred_denoised)
float_to_wav(output.wav_orig, os.path.join(
dest_dir, "original.wav"), sample_rate=output.orig_sr)
float_to_wav(output.wav_inferred_denoised, os.path.join(
dest_dir, "inferred_denoised.wav"), sample_rate=output.inferred_sr)
float_to_wav(output.wav_inferred, os.path.join(
dest_dir, "inferred.wav"), sample_rate=output.inferred_sr)
stack_images_vertically(
list_im=[
os.path.join(dest_dir, "original.png"),
os.path.join(dest_dir, "inferred_denoised.png"),
os.path.join(dest_dir, "diff.png"),
],
out_path=os.path.join(dest_dir, "comparison.png")
)
def save_val_wav(val_dir: str, sampling_rate: int, wav) -> str:
path = get_val_wav_path(val_dir)
float_to_wav(wav, path, sample_rate=sampling_rate)
return path
def save_infer_wav(infer_dir: str, sampling_rate: int, wav: np.ndarray):
path = os.path.join(infer_dir, f"{get_parent_dirname(infer_dir)}.wav")
float_to_wav(wav, path, sample_rate=sampling_rate)
def _infer(infer_dir: str, run_name: str, checkpoint_path: str,
mel_entries: List[InferMelEntry], sigma: float,
denoiser_strength: float, sentence_pause_s: Optional[float],
custom_hparams: Optional[Dict[str, str]], no_concatenation: bool,
seed: int, copy_wav_info_to: Optional[str]):
logger = prepare_logger(os.path.join(infer_dir, "log.txt"))
checkpoint = CheckpointWaveglow.load(checkpoint_path, logger)
concatenate = not no_concatenation
denoised_audio_wav_paths: List[Dict[str, Any]] = []
save_callback = partial(save_results,
infer_dir=infer_dir,
denoised_audio_wav_paths=denoised_audio_wav_paths)
inference_results, complete = infer_core(
mel_entries=mel_entries,
checkpoint=checkpoint,
custom_hparams=custom_hparams,
denoiser_strength=denoiser_strength,
sigma=sigma,
sentence_pause_s=sentence_pause_s,
logger=logger,
save_callback=save_callback,
concatenate=concatenate,
seed=seed,
)
if concatenate:
complete_wav_denoised, complete_wav_denoised_sr = complete
assert complete_wav_denoised is not None
assert complete_wav_denoised_sr is not None
float_to_wav(complete_wav_denoised,
os.path.join(infer_dir, "complete_denoised.wav"),
sample_rate=complete_wav_denoised_sr)
logger.info("Creating mel_inferred_denoised_v.png")
mel_inferred_denoised_v_plot(infer_dir, inference_results)
logger.info("Creating mel_inferred_denoised_h.png")
mel_inferred_denoised_h_plot(infer_dir, inference_results)
logger.info("Creating total.csv")
save_stats(infer_dir, inference_results)
wav_paths_json = save_denoised_audio_wav_paths(
infer_dir=infer_dir,
name=run_name,
denoised_audio_wav_paths=denoised_audio_wav_paths,
)
logger.info(
"Wrote all inferred mel paths including sampling rate into these file(s):"
)
logger.info(wav_paths_json)
if copy_wav_info_to is not None:
create_parent_folder(copy_wav_info_to)
copyfile(wav_paths_json, copy_wav_info_to)
logger.info(copy_wav_info_to)
logger.info(f"Saved output to: {infer_dir}")