Skip to content

Instantly share code, notes, and snippets.

@fnx4

fnx4/silero_tts.py

Last active May 7, 2023 08:19
Show Gist options
  • Save fnx4/cc445c38259853c93cac24abee10c1ed to your computer and use it in GitHub Desktop.
Save fnx4/cc445c38259853c93cac24abee10c1ed to your computer and use it in GitHub Desktop.
Download ZIP
OUTDATED moved to https://github.com/fnx4/silero-tts | Silero TTS: ru txt file(s)/fb2 to WAVs/opus
Raw
silero_tts.py
#!/usr/bin/env python
import os
import re
import argparse
import concurrent
import subprocess
import uuid
import torch # cuda: 1.13.1+cu117
import nltk
import transliterate
import num2words
import pydub
import tqdm
import pypandoc
###################################################
# repo: https://github.com/snakers4/silero-models #
# commit a7e61e1c6d69e981d2b8da39194ed7512ad4c3c2 #
###################################################
wrn = []
parser = argparse.ArgumentParser(description="tts", formatter_class=argparse.ArgumentDefaultsHelpFormatter,
epilog="Example: ./silero_tts.py -i folder/or/file.txt -o path/to/result -t 16 -s xenia -d cuda -r 48000 --merge")
def open_file(source, out_folder):
if not os.path.exists(out_folder):
os.makedirs(out_folder)
file = open(source, "r", encoding="utf-8")
lines = file.read().splitlines()
file.close()
print()
print(out_folder)
tts(lines, out_folder)
def experimental_svc(input_wav_blob): # UNSTABLE
svcg_path = "B:\\apps\\tts\\_other\\so-vits-svc-fork\\venv\\Scripts\\svc.exe"
model_name = "ru-saya" # https://huggingface.co/fnx/so-vits-svc-4.0-ru-saya/tree/main
checkpoint_name = "G_1500.pth"
svc_path = os.path.join(os.path.abspath(os.getcwd()), "_svc")
in_path = os.path.join(svc_path, "tmp", "in")
out_path = os.path.join(svc_path, "tmp", "out")
os.makedirs(in_path, exist_ok=True)
os.makedirs(out_path, exist_ok=True)
file_name = str(uuid.uuid4()) + ".wav"
in_file_path = os.path.join(in_path, file_name)
input_wav_blob.export(in_file_path, format="wav")
svcg_params = " infer " \
"-o " + os.path.join(out_path, file_name) + " " \
"-s " + model_name + " " \
"-m " + os.path.join(svc_path, "models", model_name, checkpoint_name) + " " \
"-c " + os.path.join(svc_path, "models", model_name, "config.json") + " " \
"-t 2 " \
"-db -25 " \
"--no-auto-predict-f0 " \
"-d cuda " \
"" + os.path.join(in_path, file_name)
subprocess.run(svcg_path + svcg_params, stdout=subprocess.PIPE, shell=True)
audio = pydub.AudioSegment.from_wav(os.path.join(out_path, file_name))
os.remove(os.path.join(in_path, file_name))
os.remove(os.path.join(out_path, file_name))
return audio
def enc_merge(merge_object):
silence_wav = pydub.AudioSegment.silent(150)
for path, dirs, files in os.walk(merge_object["out_folder"]):
combined_wav = pydub.AudioSegment.silent(0)
for file in sorted(files):
if str(file).lower().endswith(".wav"):
combined_wav = combined_wav + pydub.AudioSegment.from_wav(os.path.join(merge_object["out_folder"], file)) + silence_wav
#combined_wav = experimental_svc(combined_wav) # Testing
combined_wav_file = os.path.join(merge_object["out_folder"], "compressed_output.opus")
#if not os.path.exists(combined_wav_file):
combined_wav.export(combined_wav_file, bitrate="32k", format="opus", codec="libopus")
os.replace(os.path.join(merge_object["out_folder"], "compressed_output.opus"), os.path.join(root_out_folder, merge_object["out_file"] + ".opus"))
def enc_merge_exec(merge_objects):
print()
print("Merging and encoding, please wait... (" + cfg["threads"] + " threads)")
pbar = tqdm.tqdm(total=len(merge_objects), desc="MERGING", unit="chapter")
with concurrent.futures.ThreadPoolExecutor(max_workers=int(cfg["threads"])) as executor:
futures = [executor.submit(enc_merge, merge_object) for merge_object in merge_objects]
for future in concurrent.futures.as_completed(futures):
pbar.update(1)
def open_fb2(source):
print("converting file " + source + "...")
md = pypandoc.convert_file(source, "md")
print("extracting chapters...")
chapters = []
chapter_text = source
for line in md.splitlines():
if line == "::: section":
chapters.append(chapter_text)
chapter_text = ""
else:
# chapter_text += pypandoc.convert_text(line, "plain", format="md") # too slow
chapter_text += line + "\n"
if line == "":
chapter_text += "\r\n"
chapters.append(chapter_text)
print()
merge_objects = []
for line_num, chapter in enumerate(chapters):
chapter_name = str(chapter).partition("\n")[0]
chapter_name = pypandoc.convert_text(chapter_name, "plain", format="md")
chapter_name = chapter_name.replace(" ", "_").replace("\r", "").replace("\n", "")
chapter_name = re.sub("[^A-Za-z0-9А-Яа-яЁё_!#%№-]+", "", chapter_name.strip())
out_file = (str(line_num).zfill(6) + "_" + chapter_name).replace(".fb2", "")
out_folder = os.path.join(root_out_folder, out_file).replace(" ", "_")
merge_object = {
"out_file": out_file,
"out_folder": out_folder,
}
merge_objects.append(merge_object)
chapter = str(chapter).replace("\n", " ")
chapter = re.sub("[^A-Za-z0-9А-Яа-яЁё_\s .,;!№$&?–—-]+", "", chapter.strip())
if not os.path.exists(out_folder):
os.makedirs(out_folder)
print()
print(out_folder)
tts(chapter.splitlines(), out_folder)
if merge:
enc_merge_exec(merge_objects)
def tts(lines, out_folder):
for line_num, text in enumerate(tqdm.tqdm(lines, desc=" TTS")):
text = str(text).replace("…", ",")
text = str(text).replace("+", " плюс ")
text = str(text).replace("%", " процент ")
text = re.sub("[^A-Za-z0-9А-Яа-яЁё_\s .,;!№$&?–—-]+", "", text.strip())
t_sentences = nltk.sent_tokenize(text)
sentences = []
for sentence in t_sentences:
sentences += re.findall('.{1,850}', sentence)
# print(sentences)
for sentence_num, sentence in enumerate(sentences):
file_name = os.path.join(out_folder, "tts_" + str(line_num).zfill(6) + "_" + str(sentence_num).zfill(3) + ".wav")
sentence = re.sub(r"(\d+)", lambda x: num2words.num2words(int(x.group(0)), lang="ru"), sentence)
# print(sentence)
if text.strip() != "" and not os.path.exists(file_name):
if (re.search('[A-Za-z0-9А-Яа-яЁё]', sentence)) is not None:
sentence = transliterate.translit(sentence, language_code="ru") # TODO langs
# ssml_text = "<speak><s>" + sentence + "</s><break time='150ms'/></speak>" # sounds ugly
try:
model.save_wav(text=sentence, speaker=speaker, sample_rate=sample_rate, audio_path=file_name)
except Exception as e:
if str(e) == "Model couldn't generate your text, probably it's too long":
print("Broken tokenization EX!")
wrn_text = out_folder + " (" + str(line_num) + ") " + text + " -> " + sentence
print(wrn_text)
wrn.append(wrn_text)
sentence = ''.join(filter(str.isalpha, sentence))
model.save_wav(text=sentence, speaker=speaker, sample_rate=sample_rate, audio_path=file_name)
else:
print(str(e))
exit(1)
if __name__ == "__main__":
parser.add_argument("-i", "--input", action="store", help="input file or folder (all txt files or chapters)", required=True)
parser.add_argument("-o", "--output", action="store", help="relative output folder", default="result")
parser.add_argument("-t", "--threads", action="store", help="thread count (torch.set_num_threads value)", default="4")
parser.add_argument("-s", "--speaker", action="store", help="model speaker", default="xenia",
choices=["aidar", "baya", "kseniya", "xenia", "eugene"])
parser.add_argument("-d", "--device", action="store", help="torch.device value", default="cpu",
choices=["cpu", "cuda", "xpu", "opengl", "opencl", "ideep", "vulkan", "hpu"])
parser.add_argument("-r", "--rate", action="store", help="sample rate", default="48000")
parser.add_argument("-m", "--merge", action="store_true", help="merge wav files and save as opus")
args = parser.parse_args()
cfg = vars(args)
root = os.getcwd()
print(cfg)
nltk.download("punkt") # if needed
torch._C._jit_set_profiling_mode(False)
device = torch.device(cfg["device"])
torch.set_num_threads(int(cfg["threads"]))
sample_rate = int(cfg["rate"])
merge = cfg["merge"]
local_file = "v3_1_ru.pt" # ru_v3.pt, v3_1_ru.pt
speaker = cfg["speaker"]
if not os.path.isfile(local_file):
torch.hub.download_url_to_file("https://models.silero.ai/models/tts/ru/" + local_file, local_file)
model = torch.package.PackageImporter(local_file).load_pickle("tts_models", "model")
model.to(device)
root_out_folder = cfg["output"]
source = cfg["input"]
if os.path.isdir(source):
merge_objects = []
for path, dirs, files in os.walk(source):
for file_name in sorted(files):
out_file = str(file_name).replace(" ", "_").replace(".txt", "")
out_file = re.sub("[^A-Za-z0-9А-Яа-яЁё_!#%№-]+", "", out_file.strip())
out_folder = os.path.join(root_out_folder, out_file)
open_file(os.path.join(path, file_name), out_folder)
merge_object = {
"out_file": out_file,
"out_folder": out_folder,
}
merge_objects.append(merge_object)
if merge:
enc_merge_exec(merge_objects)
elif os.path.isfile(source):
if str(source).lower().endswith(".fb2"):
open_fb2(source)
else:
print(source)
print(root_out_folder)
open_file(source, root_out_folder)
if merge:
merge_object = {
"out_file": "compressed_output.opus",
"out_folder": root_out_folder,
}
print("Merging and encoding, please wait...")
enc_merge(merge_object)
else:
print("Error: input file or folder not found")
exit(1)
print("\nWarnings: ")
print(wrn)
print("\nFinished!")
exit(0)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment