Spaces:
Sleeping
Sleeping
File size: 9,733 Bytes
8aa346c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 |
import numpy as np
import gradio as gr
import torch
import os
import warnings
from gradio.processing_utils import convert_to_16_bit_wav
from typing import Dict, List, Optional, Union
import utils
from infer import get_net_g, infer
from models import SynthesizerTrn
from models_jp_extra import SynthesizerTrn as SynthesizerTrnJPExtra
from .log import logger
from .constants import (
DEFAULT_ASSIST_TEXT_WEIGHT,
DEFAULT_LENGTH,
DEFAULT_LINE_SPLIT,
DEFAULT_NOISE,
DEFAULT_NOISEW,
DEFAULT_SDP_RATIO,
DEFAULT_SPLIT_INTERVAL,
DEFAULT_STYLE,
DEFAULT_STYLE_WEIGHT,
)
class Model:
def __init__(
self, model_path: str, config_path: str, style_vec_path: str, device: str
):
self.model_path: str = model_path
self.config_path: str = config_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.num_styles: int = self.hps.data.num_styles
if hasattr(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)}
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)})"
)
self.style_vectors: np.ndarray = np.load(self.style_vec_path)
if self.style_vectors.shape[0] != self.num_styles:
raise ValueError(
f"The number of styles ({self.num_styles}) does not match the number of style vectors ({self.style_vectors.shape[0]})"
)
self.net_g: Union[SynthesizerTrn, SynthesizerTrnJPExtra, None] = None
def load_net_g(self):
self.net_g = get_net_g(
model_path=self.model_path,
version=self.hps.version,
device=self.device,
hps=self.hps,
)
def get_style_vector(self, style_id: int, weight: float = 1.0) -> np.ndarray:
mean = self.style_vectors[0]
style_vec = self.style_vectors[style_id]
style_vec = mean + (style_vec - mean) * weight
return style_vec
def get_style_vector_from_audio(
self, audio_path: str, weight: float = 1.0
) -> np.ndarray:
from style_gen import get_style_vector
xvec = get_style_vector(audio_path)
mean = self.style_vectors[0]
xvec = mean + (xvec - mean) * weight
return xvec
def infer(
self,
text: str,
language: str = "JP",
sid: int = 0,
reference_audio_path: Optional[str] = None,
sdp_ratio: float = DEFAULT_SDP_RATIO,
noise: float = DEFAULT_NOISE,
noisew: float = DEFAULT_NOISEW,
length: float = DEFAULT_LENGTH,
line_split: bool = DEFAULT_LINE_SPLIT,
split_interval: float = DEFAULT_SPLIT_INTERVAL,
assist_text: Optional[str] = None,
assist_text_weight: float = DEFAULT_ASSIST_TEXT_WEIGHT,
use_assist_text: bool = False,
style: str = DEFAULT_STYLE,
style_weight: float = DEFAULT_STYLE_WEIGHT,
given_tone: Optional[list[int]] = None,
) -> 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"):
raise ValueError(
"The model is trained with JP-Extra, but the language is not JP"
)
if reference_audio_path == "":
reference_audio_path = None
if assist_text == "" or not use_assist_text:
assist_text = None
if self.net_g is None:
self.load_net_g()
if reference_audio_path is None:
style_id = self.style2id[style]
style_vector = self.get_style_vector(style_id, style_weight)
else:
style_vector = self.get_style_vector_from_audio(
reference_audio_path, style_weight
)
if not line_split:
with torch.no_grad():
audio = infer(
text=text,
sdp_ratio=sdp_ratio,
noise_scale=noise,
noise_scale_w=noisew,
length_scale=length,
sid=sid,
language=language,
hps=self.hps,
net_g=self.net_g,
device=self.device,
assist_text=assist_text,
assist_text_weight=assist_text_weight,
style_vec=style_vector,
given_tone=given_tone,
)
else:
texts = text.split("\n")
texts = [t for t in texts if t != ""]
audios = []
with torch.no_grad():
for i, t in enumerate(texts):
audios.append(
infer(
text=t,
sdp_ratio=sdp_ratio,
noise_scale=noise,
noise_scale_w=noisew,
length_scale=length,
sid=sid,
language=language,
hps=self.hps,
net_g=self.net_g,
device=self.device,
assist_text=assist_text,
assist_text_weight=assist_text_weight,
style_vec=style_vector,
)
)
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")
return (self.hps.data.sampling_rate, audio)
class ModelHolder:
def __init__(self, root_dir: str, device: str):
self.root_dir: str = root_dir
self.device: str = device
self.model_files_dict: Dict[str, List[str]] = {}
self.current_model: Optional[Model] = None
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_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")
]
if len(model_files) == 0:
logger.warning(
f"No model files found in {self.root_dir}/{model_name}, so skip it"
)
continue
self.model_files_dict[model_name] = model_files
self.model_names.append(model_name)
def load_model_gr(
self, model_name: str, model_path: str
) -> tuple[gr.Dropdown, gr.Button, gr.Dropdown]:
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 not None
and self.current_model.model_path == model_path
):
# Already loaded
speakers = list(self.current_model.spk2id.keys())
styles = list(self.current_model.style2id.keys())
return (
gr.Dropdown(choices=styles, value=styles[0]),
gr.Button(interactive=True, value="音声合成"),
gr.Dropdown(choices=speakers, value=speakers[0]),
)
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"),
device=self.device,
)
speakers = list(self.current_model.spk2id.keys())
styles = list(self.current_model.style2id.keys())
return (
gr.Dropdown(choices=styles, value=styles[0]),
gr.Button(interactive=True, value="音声合成"),
gr.Dropdown(choices=speakers, value=speakers[0]),
)
def update_model_files_gr(self, model_name: str) -> gr.Dropdown:
model_files = self.model_files_dict[model_name]
return gr.Dropdown(choices=model_files, value=model_files[0])
def update_model_names_gr(self) -> tuple[gr.Dropdown, gr.Dropdown, gr.Button]:
self.refresh()
initial_model_name = self.model_names[0]
initial_model_files = self.model_files_dict[initial_model_name]
return (
gr.Dropdown(choices=self.model_names, value=initial_model_name),
gr.Dropdown(choices=initial_model_files, value=initial_model_files[0]),
gr.Button(interactive=False), # For tts_button
)
|