Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,47 +1,57 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
-
from transformers import
|
| 3 |
import soundfile as sf
|
| 4 |
import torch
|
| 5 |
import os
|
| 6 |
|
| 7 |
-
#
|
| 8 |
-
model_name = "Soul-AILab/SoulX-Podcast-1.7B"
|
| 9 |
-
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 10 |
-
model = AutoModel.from_pretrained(
|
| 11 |
-
model_name,
|
| 12 |
-
dtype=torch.float16, # 替换 deprecated 的 torch_dtype
|
| 13 |
-
# 移除 device_map="auto",改用手动分配设备(兼容无accelerate环境)
|
| 14 |
-
)
|
| 15 |
-
# 手动将模型移到GPU(无GPU自动用CPU)
|
| 16 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 17 |
-
model = model.to(device)
|
| 18 |
|
| 19 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
def generate_speech(text):
|
| 21 |
if not text.strip():
|
| 22 |
return None, "错误:请输入有效文本!"
|
| 23 |
|
| 24 |
-
#
|
| 25 |
-
|
| 26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
with torch.no_grad():
|
| 28 |
-
audio_output = model.generate(**inputs)
|
| 29 |
|
| 30 |
# 保存音频
|
| 31 |
output_path = "output.wav"
|
| 32 |
-
sf.write(output_path, audio_output[0].
|
| 33 |
|
| 34 |
-
return output_path, "
|
| 35 |
|
| 36 |
-
#
|
| 37 |
-
with gr.Blocks(title="
|
| 38 |
-
gr.Markdown("# 🎤
|
| 39 |
-
gr.Markdown("
|
| 40 |
|
| 41 |
with gr.Row():
|
| 42 |
text_input = gr.Textbox(
|
| 43 |
label="输入文本",
|
| 44 |
-
placeholder="
|
| 45 |
lines=5
|
| 46 |
)
|
| 47 |
audio_output = gr.Audio(label="生成的语音", type="filepath")
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
+
from transformers import AutoModelForTextToSpeech, AutoTokenizer, pipeline
|
| 3 |
import soundfile as sf
|
| 4 |
import torch
|
| 5 |
import os
|
| 6 |
|
| 7 |
+
# 初始化中英双语TTS管道(轻量模型,总体积<5GB)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
| 9 |
|
| 10 |
+
# 英文TTS(fastspeech2,体积~2GB)
|
| 11 |
+
en_tokenizer = AutoTokenizer.from_pretrained("facebook/fastspeech2-en-ljspeech")
|
| 12 |
+
en_model = AutoModelForTextToSpeech.from_pretrained("facebook/fastspeech2-en-ljspeech").to(device)
|
| 13 |
+
|
| 14 |
+
# 中文TTS(Chinese-FastSpeech2,体积~3GB)
|
| 15 |
+
zh_tokenizer = AutoTokenizer.from_pretrained("bakerk1234/Chinese-FastSpeech2")
|
| 16 |
+
zh_model = AutoModelForTextToSpeech.from_pretrained("bakerk1234/Chinese-FastSpeech2").to(device)
|
| 17 |
+
|
| 18 |
+
# 语音生成函数(自动识别语言,切换模型)
|
| 19 |
def generate_speech(text):
|
| 20 |
if not text.strip():
|
| 21 |
return None, "错误:请输入有效文本!"
|
| 22 |
|
| 23 |
+
# 简单语言识别(中文含中文字符,英文不含)
|
| 24 |
+
is_chinese = any('\u4e00' <= char <= '\u9fff' for char in text)
|
| 25 |
|
| 26 |
+
if is_chinese:
|
| 27 |
+
tokenizer = zh_tokenizer
|
| 28 |
+
model = zh_model
|
| 29 |
+
samplerate = 22050 # 中文模型采样率
|
| 30 |
+
else:
|
| 31 |
+
tokenizer = en_tokenizer
|
| 32 |
+
model = en_model
|
| 33 |
+
samplerate = 22050 # 英文模型采样率
|
| 34 |
+
|
| 35 |
+
# 文本编码+生成语音
|
| 36 |
+
inputs = tokenizer(text, return_tensors="pt").to(device)
|
| 37 |
with torch.no_grad():
|
| 38 |
+
audio_output = model.generate(**inputs).cpu().numpy()
|
| 39 |
|
| 40 |
# 保存音频
|
| 41 |
output_path = "output.wav"
|
| 42 |
+
sf.write(output_path, audio_output[0].T, samplerate=samplerate) # 调整维度适配保存
|
| 43 |
|
| 44 |
+
return output_path, f"语音生成成功!(使用{'中文' if is_chinese else '英文'}轻量模型)"
|
| 45 |
|
| 46 |
+
# 界面保持不变
|
| 47 |
+
with gr.Blocks(title="中英双语TTS(轻量版)") as demo:
|
| 48 |
+
gr.Markdown("# 🎤 轻量中英双语文本转语音")
|
| 49 |
+
gr.Markdown("基于FastSpeech2模型,体积小(<5GB),适配免费Space,支持中英双语输入")
|
| 50 |
|
| 51 |
with gr.Row():
|
| 52 |
text_input = gr.Textbox(
|
| 53 |
label="输入文本",
|
| 54 |
+
placeholder="请输入中文或英文文本(建议≤300字)...",
|
| 55 |
lines=5
|
| 56 |
)
|
| 57 |
audio_output = gr.Audio(label="生成的语音", type="filepath")
|