|
from io import BytesIO |
|
import requests |
|
from os.path import exists, join |
|
from espnet2.bin.tts_inference import Text2Speech |
|
from enum import Enum |
|
from .formatter import preprocess_text |
|
from .stress import sentence_to_stress, stress_dict, stress_with_model |
|
from torch import no_grad |
|
import numpy as np |
|
import time |
|
import soundfile as sf |
|
|
|
|
|
class Voices(Enum): |
|
"""List of available voices for the model.""" |
|
|
|
Olena = 4 |
|
Mykyta = 3 |
|
Lada = 2 |
|
Dmytro = 1 |
|
Olga = 5 |
|
|
|
|
|
class Stress(Enum): |
|
"""Options how to stress sentence. |
|
- `dictionary` - performs lookup in dictionary, taking into account grammatical case of a word and its' neighbors |
|
- `model` - stress using transformer model""" |
|
|
|
Dictionary = "dictionary" |
|
Model = "model" |
|
|
|
|
|
class TTS: |
|
""" """ |
|
|
|
def __init__(self, cache_folder=None, device="cpu") -> None: |
|
""" |
|
Class to setup a text-to-speech engine, from download to model creation. \n |
|
Downloads or uses files from `cache_folder` directory. \n |
|
By default stores in current directory.""" |
|
self.device = device |
|
self.__setup_cache(cache_folder) |
|
|
|
def tts(self, text: str, voice: int, stress: str, output_fp=BytesIO(), speed=1.0): |
|
""" |
|
Run a Text-to-Speech engine and output to `output_fp` BytesIO-like object. |
|
- `text` - your model input text. |
|
- `voice` - one of predefined voices from `Voices` enum. |
|
- `stress` - stress method options, predefined in `Stress` enum. |
|
- `output_fp` - file-like object output. Stores in RAM by default. |
|
""" |
|
|
|
if stress not in [option.value for option in Stress]: |
|
raise ValueError( |
|
f"Invalid value for stress option selected! Please use one of the following values: {', '.join([option.value for option in Stress])}." |
|
) |
|
|
|
if stress == Stress.Model.value: |
|
stress = True |
|
else: |
|
stress = False |
|
if voice not in [option.value for option in Voices]: |
|
raise ValueError( |
|
f"Invalid value for voice selected! Please use one of the following values: {', '.join([option.value for option in Voices])}." |
|
) |
|
|
|
text = preprocess_text(text) |
|
text = sentence_to_stress(text, stress_with_model if stress else stress_dict) |
|
|
|
|
|
with no_grad(): |
|
start = time.time() |
|
wav = self.synthesizer( |
|
text, sids=np.array(voice), decode_conf={"alpha": 1 / speed} |
|
)["wav"] |
|
|
|
rtf = (time.time() - start) / (len(wav) / self.synthesizer.fs) |
|
print(f"RTF = {rtf:5f}") |
|
|
|
sf.write( |
|
output_fp, |
|
wav.view(-1).cpu().numpy(), |
|
self.synthesizer.fs, |
|
"PCM_16", |
|
format="wav", |
|
) |
|
|
|
output_fp.seek(0) |
|
|
|
return output_fp, text |
|
|
|
def __setup_cache(self, cache_folder=None): |
|
"""Downloads models and stores them into `cache_folder`. By default stores in current directory.""" |
|
print("downloading uk/mykyta/vits-tts") |
|
release_number = "v4.0.0" |
|
model_link = f"https://github.com/robinhad/ukrainian-tts/releases/download/{release_number}/model.pth" |
|
config_link = f"https://github.com/robinhad/ukrainian-tts/releases/download/{release_number}/config.yaml" |
|
|
|
if cache_folder is None: |
|
cache_folder = "." |
|
|
|
model_path = join(cache_folder, "model.pth") |
|
config_path = join(cache_folder, "config.yaml") |
|
|
|
self.__download(model_link, model_path) |
|
self.__download(config_link, config_path) |
|
|
|
self.synthesizer = Text2Speech( |
|
train_config="config.yaml", |
|
model_file="model.pth", |
|
device=self.device, |
|
|
|
noise_scale=0.333, |
|
noise_scale_dur=0.333, |
|
) |
|
|
|
def __download(self, url, file_name): |
|
"""Downloads file from `url` into local `file_name` file.""" |
|
if not exists(file_name): |
|
print(f"Downloading {file_name}") |
|
r = requests.get(url, allow_redirects=True) |
|
with open(file_name, "wb") as file: |
|
file.write(r.content) |
|
else: |
|
print(f"Found {file_name}. Skipping download...") |
|
|
