tuna2134/sbv2-v2
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Feat: add editor server backend

Browse Source
This commit is contained in:
litagin02
2024-02-23 13:21:14 +09:00
parent de539e74f9
commit e16cdad6e9
10 changed files with 423 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
common/constants.py
View File

@@ -5,6 +5,10 @@ import enum
GRADIO_THEME: str = "NoCrypt/miku"
LATEST_VERSION: str = "2.3"
USER_DICT_CSV_PATH: str = "dict/user_dict.csv"
USER_DICT_PATH: str = "dict/user.dic"
DEFAULT_STYLE: str = "Neutral"
DEFAULT_STYLE_WEIGHT: float = 5.0

110
common/tts_model.py
View File

@@ -1,6 +1,7 @@
import os
import warnings
from typing import Dict, List, Optional, Union
from pathlib import Path
from typing import Optional, Union
import gradio as gr
import numpy as np
@@ -29,12 +30,13 @@ from .log import logger
def adjust_voice(fs, wave, pitch_scale, intonation_scale):
if pitch_scale == 1.0 and intonation_scale == 1.0:
# 初期値の場合は、音質劣化を避けるためにそのまま返す
return fs, wave
# pyworldでf0を加工して合成
# pyworldよりもよいのがあるかもしれないが……
wave = wave.astype(np.double)
logger.debug(f"wave: shape={wave.shape}, max={max(wave)}, min={min(wave)}")
f0, t = pyworld.harvest(wave, fs)
# 質が高そうだしとりあえずharvestにしておく
@@ -43,7 +45,6 @@ def adjust_voice(fs, wave, pitch_scale, intonation_scale):
non_zero_f0 = [f for f in f0 if f != 0]
f0_mean = sum(non_zero_f0) / len(non_zero_f0)
logger.debug(f"f0: shape={f0.shape}, mean={f0_mean}, max={max(f0)}")
for i, f in enumerate(f0):
if f == 0:
@@ -56,21 +57,21 @@ def adjust_voice(fs, wave, pitch_scale, intonation_scale):
class Model:
def __init__(
self, model_path: str, config_path: str, style_vec_path: str, device: str
self, model_path: Path, config_path: Path, style_vec_path: Path, device: str
):
self.model_path: str = model_path
self.config_path: str = config_path
self.model_path: Path = model_path
self.config_path: Path = config_path
self.style_vec_path: Path = style_vec_path
self.device: str = device
self.style_vec_path: str = style_vec_path
self.hps: utils.HParams = utils.get_hparams_from_file(self.config_path)
self.spk2id: Dict[str, int] = self.hps.data.spk2id
self.id2spk: Dict[int, str] = {v: k for k, v in self.spk2id.items()}
self.spk2id: dict[str, int] = self.hps.data.spk2id
self.id2spk: dict[int, str] = {v: k for k, v in self.spk2id.items()}
self.num_styles: int = self.hps.data.num_styles
if hasattr(self.hps.data, "style2id"):
self.style2id: Dict[str, int] = self.hps.data.style2id
self.style2id: dict[str, int] = self.hps.data.style2id
else:
self.style2id: Dict[str, int] = {str(i): i for i in range(self.num_styles)}
self.style2id: dict[str, int] = {str(i): i for i in range(self.num_styles)}
if len(self.style2id) != self.num_styles:
raise ValueError(
f"Number of styles ({self.num_styles}) does not match the number of style2id ({len(self.style2id)})"
@@ -86,7 +87,7 @@ class Model:
def load_net_g(self):
self.net_g = get_net_g(
model_path=self.model_path,
model_path=str(self.model_path),
version=self.hps.version,
device=self.device,
hps=self.hps,
@@ -128,6 +129,7 @@ class Model:
given_tone: Optional[list[int]] = None,
pitch_scale: float = 1.0,
intonation_scale: float = 1.0,
ignore_unknown: bool = False,
) -> tuple[int, np.ndarray]:
logger.info(f"Start generating audio data from text:\n{text}")
if language != "JP" and self.hps.version.endswith("JP-Extra"):
@@ -165,6 +167,7 @@ class Model:
assist_text_weight=assist_text_weight,
style_vec=style_vector,
given_tone=given_tone,
ignore_unknown=ignore_unknown,
)
else:
texts = text.split("\n")
@@ -187,14 +190,12 @@ class Model:
assist_text=assist_text,
assist_text_weight=assist_text_weight,
style_vec=style_vector,
ignore_unknown=ignore_unknown,
)
)
if i != len(texts) - 1:
audios.append(np.zeros(int(44100 * split_interval)))
audio = np.concatenate(audios)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
audio = convert_to_16_bit_wav(audio)
logger.info("Audio data generated successfully")
fs, audio = adjust_voice(
fs=self.hps.data.sampling_rate,
@@ -202,46 +203,79 @@ class Model:
pitch_scale=pitch_scale,
intonation_scale=intonation_scale,
)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
audio = convert_to_16_bit_wav(audio)
return (fs, audio)
class ModelHolder:
def __init__(self, root_dir: str, device: str):
self.root_dir: str = root_dir
def __init__(self, root_dir: Path, device: str):
self.root_dir: Path = root_dir
self.device: str = device
self.model_files_dict: Dict[str, List[str]] = {}
self.model_files_dict: dict[str, list[Path]] = {}
self.current_model: Optional[Model] = None
self.model_names: List[str] = []
self.models: List[Model] = []
self.model_names: list[str] = []
self.models: list[Model] = []
self.refresh()
def refresh(self):
self.model_files_dict = {}
self.model_names = []
self.current_model = None
model_dirs = [
d
for d in os.listdir(self.root_dir)
if os.path.isdir(os.path.join(self.root_dir, d))
]
for model_name in model_dirs:
model_dir = os.path.join(self.root_dir, model_name)
model_dirs = [d for d in self.root_dir.iterdir() if d.is_dir()]
for model_dir in model_dirs:
model_files = [
os.path.join(model_dir, f)
for f in os.listdir(model_dir)
if f.endswith(".pth") or f.endswith(".pt") or f.endswith(".safetensors")
f
for f in model_dir.iterdir()
if f.suffix in [".pth", ".pt", ".safetensors"]
]
if len(model_files) == 0:
logger.warning(
f"No model files found in {self.root_dir}/{model_name}, so skip it"
)
logger.warning(f"No model files found in {model_dir}, so skip it")
continue
self.model_files_dict[model_name] = model_files
self.model_names.append(model_name)
self.model_files_dict[model_dir.name] = model_files
self.model_names.append(model_dir.name)
def models_info(self):
if hasattr(self, "_models_info"):
return self._models_info
result = []
for name, files in self.model_files_dict.items():
# Get styles
config_path = self.root_dir / name / "config.json"
hps = utils.get_hparams_from_file(config_path)
style2id: dict[str, int] = hps.data.style2id
styles = list(style2id.keys())
result.append(
{
"name": name,
"files": [str(f) for f in files],
"styles": styles,
}
)
self._models_info = result
return result
def load_model(self, model_name: str, model_path_str: str):
model_path = Path(model_path_str)
if model_name not in self.model_files_dict:
raise ValueError(f"Model `{model_name}` is not found")
if model_path not in self.model_files_dict[model_name]:
raise ValueError(f"Model file `{model_path}` is not found")
if self.current_model is None or self.current_model.model_path != model_path:
self.current_model = Model(
model_path=model_path,
config_path=self.root_dir / model_name / "config.json",
style_vec_path=self.root_dir / model_name / "style_vectors.npy",
device=self.device,
)
return self.current_model
def load_model_gr(
self, model_name: str, model_path: str
self, model_name: str, model_path_str: str
) -> tuple[gr.Dropdown, gr.Button, gr.Dropdown]:
model_path = Path(model_path_str)
if model_name not in self.model_files_dict:
raise ValueError(f"Model `{model_name}` is not found")
if model_path not in self.model_files_dict[model_name]:
@@ -260,8 +294,8 @@ class ModelHolder:
)
self.current_model = Model(
model_path=model_path,
config_path=os.path.join(self.root_dir, model_name, "config.json"),
style_vec_path=os.path.join(self.root_dir, model_name, "style_vectors.npy"),
config_path=self.root_dir / model_name / "config.json",
style_vec_path=self.root_dir / model_name / "style_vectors.npy",
device=self.device,
)
speakers = list(self.current_model.spk2id.keys())