Li Wei Chen
fix: correct G2P duration estimation and remove punctuation spaces
80cdf12
raw
history blame contribute delete
12 kB
from __future__ import annotations
import logging
import os
import re
from dataclasses import dataclass
from typing import Any
import gradio as gr
import numpy as np
try:
import spaces
USING_SPACES = True
except ImportError:
USING_SPACES = False
MODEL_ID = "formospeech/omnivoice-taiwanese-hakka"
DIALECT_LABELS = [
"客語四縣腔",
"客語海陸腔",
"客語大埔腔",
"客語饒平腔",
"客語詔安腔",
"客語南四縣腔",
]
DIALECT_TO_LANG_GROUP = {
"客語四縣腔": "hak_sx",
"客語海陸腔": "hak_hl",
"客語大埔腔": "hak_dp",
"客語饒平腔": "hak_rp",
"客語詔安腔": "hak_za",
"客語南四縣腔": "hak_nsx",
}
DEFAULT_SPEED = 1.0
DEFAULT_STEPS = 32
EXAMPLES = [
[
"客語四縣腔",
"食飯愛正經食,正毋會食到半出半入。",
"refs/0000001_0.15-0.93.wav",
"恁早。",
DEFAULT_SPEED,
DEFAULT_STEPS,
False,
],
[
"客語四縣腔",
"食飯愛正經食,正毋會食到半出半入。",
"refs/0000002_0.15-2.73.wav",
"你今晡日著到恁派頭。",
DEFAULT_SPEED,
DEFAULT_STEPS,
False,
],
[
"客語四縣腔",
"歸條路吊等長長个花燈,祈求風調雨順,歸屋下人个心願,親像花燈下燒暖个光華。",
"refs/0000002_0.15-2.73.wav",
"你今晡日著到恁派頭。",
DEFAULT_SPEED,
DEFAULT_STEPS,
False,
],
]
@dataclass
class RuntimeState:
model: Any | None
generation_config_cls: Any | None
sampling_rate: int | None
device: str
dtype_name: str
load_error: str | None = None
def gpu_decorator(func):
if USING_SPACES:
return spaces.GPU(func)
return func
def get_best_device() -> str:
try:
import torch
except Exception:
return "cpu"
if torch.cuda.is_available():
return "cuda"
if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available():
return "mps"
return "cpu"
def load_runtime() -> RuntimeState:
device = get_best_device()
dtype_name = "float16" if device == "cuda" else "float32"
try:
import torch
from omnivoice import OmniVoice, OmniVoiceGenerationConfig
except Exception as exc:
return RuntimeState(
model=None,
generation_config_cls=None,
sampling_rate=None,
device=device,
dtype_name=dtype_name,
load_error=f"依賴載入失敗:{type(exc).__name__}: {exc}",
)
dtype = torch.float16 if device == "cuda" else torch.float32
try:
logging.info("Loading model %s on %s with %s", MODEL_ID, device, dtype_name)
model = OmniVoice.from_pretrained(
MODEL_ID,
device_map=device,
dtype=dtype,
load_asr=False,
)
except Exception as exc:
return RuntimeState(
model=None,
generation_config_cls=OmniVoiceGenerationConfig,
sampling_rate=None,
device=device,
dtype_name=dtype_name,
load_error=f"模型載入失敗:{type(exc).__name__}: {exc}",
)
return RuntimeState(
model=model,
generation_config_cls=OmniVoiceGenerationConfig,
sampling_rate=model.sampling_rate,
device=device,
dtype_name=dtype_name,
)
RUNTIME = load_runtime()
def startup_status() -> str:
if RUNTIME.load_error:
return RUNTIME.load_error
return (
f"模型已載入:{MODEL_ID}\n"
f"裝置:{RUNTIME.device}\n"
f"推論精度:{RUNTIME.dtype_name}"
)
def apply_g2p(text: str, dialect: str) -> str:
from formog2p.hakka.g2p import g2p
lang_group = DIALECT_TO_LANG_GROUP.get(dialect, "hak_sx")
result = g2p(text, lang_group=lang_group, pronunciation_type="pinyin")
joined = " ".join(result.pronunciations).upper()
joined = re.sub(r"\s+([,。!?;:、…「」『』【】〔〕()])", r"\1", joined)
joined = re.sub(r"([,。!?;:、…「」『』【】〔〕()])\s+", r"\1", joined)
return joined
def validate_inputs(
dialect: str | None,
text: str,
ref_audio: str | None,
ref_text: str,
) -> str | None:
if dialect not in DIALECT_LABELS:
return "請先選擇客語腔調。"
if not text or not text.strip():
return "請輸入要合成的文字。"
if not ref_audio:
return "請上傳參考音檔。"
if not ref_text or not ref_text.strip():
return "請輸入參考文本。"
return None
def to_audio_output(audio: np.ndarray, sampling_rate: int) -> tuple[int, np.ndarray]:
waveform = np.asarray(audio)
if waveform.ndim > 1:
waveform = np.squeeze(waveform)
waveform = np.clip(waveform, -1.0, 1.0)
return sampling_rate, (waveform * 32767).astype(np.int16)
@gpu_decorator
def synthesize(
dialect: str | None,
text: str,
ref_audio: str | None,
ref_text: str,
speed: float,
num_step: int,
use_g2p: bool,
) -> tuple[tuple[int, np.ndarray] | None, str]:
error = validate_inputs(dialect, text, ref_audio, ref_text)
if error:
return None, error
if (
RUNTIME.load_error
or RUNTIME.model is None
or RUNTIME.generation_config_cls is None
):
return None, startup_status()
try:
original_text = text.strip()
g2p_note = ""
duration_override = None
generation_config = RUNTIME.generation_config_cls(
num_step=int(num_step),
guidance_scale=2.0,
denoise=True,
preprocess_prompt=True,
postprocess_output=True,
)
voice_clone_prompt = RUNTIME.model.create_voice_clone_prompt(
ref_audio=ref_audio,
ref_text=ref_text.strip(),
preprocess_prompt=True,
)
if use_g2p:
input_text = apply_g2p(original_text, dialect)
g2p_note = f";G2P 轉換:{input_text}"
# Estimate duration from original Chinese text to avoid weight inflation
# caused by tone number digits (weight 3.5) in the G2P output.
num_ref_tokens = voice_clone_prompt.ref_audio_tokens.size(-1)
frame_rate = RUNTIME.model.audio_tokenizer.config.frame_rate
est_frames = RUNTIME.model.duration_estimator.estimate_duration(
original_text, voice_clone_prompt.ref_text, num_ref_tokens
)
duration_override = est_frames / float(speed) / frame_rate
else:
input_text = original_text
generate_kwargs: dict[str, Any] = {
"text": input_text,
"voice_clone_prompt": voice_clone_prompt,
"instruct": dialect,
"generation_config": generation_config,
"language": "zh",
}
if duration_override is not None:
generate_kwargs["duration"] = duration_override
elif speed != DEFAULT_SPEED:
generate_kwargs["speed"] = float(speed)
audio = RUNTIME.model.generate(**generate_kwargs)
if not audio:
return None, "模型沒有回傳音訊。"
return (
to_audio_output(audio[0], int(RUNTIME.sampling_rate or 24000)),
f"合成完成。腔調:{dialect};speed={speed:.2f};steps={int(num_step)}{g2p_note}",
)
except Exception as exc:
return None, f"合成失敗:{type(exc).__name__}: {exc}"
def build_demo() -> gr.Blocks:
with gr.Blocks(title="臺灣客語語音生成系統") as demo:
with gr.Column():
gr.Markdown(
"""
# 臺灣客語語音合成系統
### Taiwanese Hakka Text-to-Speech System
### 研發團隊
- **[李鴻欣 Hung-Shin Lee](mailto:hungshinlee@gmail.com)**
- **[陳力瑋 Li-Wei Chen](mailto:wayne900619@gmail.com)**
### 合作單位
- **[國立聯合大學智慧客家實驗室](https://www.gohakka.org)**
"""
)
with gr.Row(equal_height=False):
with gr.Column(scale=11, elem_classes="panel"):
dialect = gr.Dropdown(
choices=DIALECT_LABELS,
value=None,
allow_custom_value=False,
label="客語腔調",
info="此模型用 instruct 控制腔調,推論前必選。",
)
text = gr.Textbox(
label="要合成的文字",
lines=4,
placeholder="例如:這下來試看啊,客語語音合成聽起來仰般。",
)
ref_audio = gr.Audio(
label="參考音檔",
type="filepath",
)
ref_text = gr.Textbox(
label="參考文本",
lines=2,
placeholder="請填寫參考音檔對應的逐字文本。",
)
use_g2p = gr.Checkbox(
value=False,
label="使用 G2P 轉換",
info="勾選後會先用 formog2p 將漢字轉成拼音(大寫)再輸入模型;不勾選則直接輸入原文。",
)
with gr.Accordion("進階設定", open=False):
speed = gr.Slider(
minimum=0.5,
maximum=1.5,
value=DEFAULT_SPEED,
step=0.05,
label="Speed",
info="1.0 為預設語速;越大越快。",
)
num_step = gr.Slider(
minimum=4,
maximum=32,
value=DEFAULT_STEPS,
step=1,
label="Inference Steps",
info="步數越高通常品質越穩,但速度較慢。",
)
submit = gr.Button("開始合成", variant="primary")
with gr.Column(scale=9):
output_audio = gr.Audio(
label="合成結果",
type="numpy",
)
status = gr.Textbox(
label="狀態",
value=startup_status(),
lines=6,
interactive=False,
)
submit.click(
fn=synthesize,
inputs=[dialect, text, ref_audio, ref_text, speed, num_step, use_g2p],
outputs=[output_audio, status],
)
gr.Examples(
examples=EXAMPLES,
inputs=[dialect, text, ref_audio, ref_text, speed, num_step, use_g2p],
label="範例",
)
return demo
demo = build_demo()
def main() -> None:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s %(levelname)s %(message)s",
)
demo.queue().launch(
css="@import url(https://tauhu.tw/tauhu-oo.css);",
theme=gr.themes.Default(
font=(
"tauhu-oo",
gr.themes.GoogleFont("Source Sans Pro"),
"ui-sans-serif",
"system-ui",
"sans-serif",
)
),
server_name="0.0.0.0",
server_port=int(os.getenv("PORT", "7860")),
)
if __name__ == "__main__":
main()