Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -9,54 +9,48 @@ import re
|
|
| 9 |
from pydub import AudioSegment
|
| 10 |
from transformers import pipeline
|
| 11 |
|
| 12 |
-
# --- НАСТРОЙКИ ФЭНТЕЗИ ---
|
| 13 |
VOICE_CONFIG = {
|
| 14 |
-
"narrator": {"voice": "ru-RU-DmitryNeural", "pitch": "-7Hz", "rate": "-5%"},
|
| 15 |
-
"male": {"voice": "ru-RU-DenisNeural", "pitch": "-2Hz", "rate": "+0%"},
|
| 16 |
-
"female": {"voice": "ru-RU-SvetlanaNeural","pitch": "+5Hz", "rate": "+5%"}
|
| 17 |
}
|
| 18 |
|
| 19 |
TEMP_DIR = tempfile.gettempdir()
|
| 20 |
|
| 21 |
# --- ЗАГРУЗКА МАЛЕНЬКОЙ НЕЙРОСЕТИ ---
|
| 22 |
-
#
|
| 23 |
MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct"
|
| 24 |
|
| 25 |
-
print(f"🚀 Загрузка
|
| 26 |
try:
|
| 27 |
-
# Создаем пайплайн для генерации текста
|
| 28 |
pipe = pipeline(
|
| 29 |
"text-generation",
|
| 30 |
model=MODEL_ID,
|
| 31 |
-
device_map="auto",
|
| 32 |
max_new_tokens=2048,
|
| 33 |
trust_remote_code=True
|
| 34 |
)
|
| 35 |
-
print("✅ Модель
|
| 36 |
except Exception as e:
|
| 37 |
print(f"❌ Ошибка загрузки модели: {e}")
|
| 38 |
pipe = None
|
| 39 |
|
| 40 |
def analyze_text_with_tiny_ai(text):
|
| 41 |
-
"""
|
| 42 |
-
Использует маленькую модель для разбора текста.
|
| 43 |
-
"""
|
| 44 |
if not pipe:
|
| 45 |
return [{"text": text, "role": "narrator"}]
|
| 46 |
|
| 47 |
-
# Простой промпт для маленькой модели.
|
| 48 |
-
# Маленькие модели любят конкретику.
|
| 49 |
system_prompt = (
|
| 50 |
-
"Ты редактор. Твоя задача -
|
| 51 |
-
"
|
| 52 |
-
"
|
| 53 |
)
|
| 54 |
|
| 55 |
-
user_prompt = f"""Разбей
|
| 56 |
-
|
| 57 |
"{text}"
|
| 58 |
|
| 59 |
-
Пример ответа:
|
| 60 |
[{{"text": "- Привет", "role": "male"}}, {{"text": "- сказала она", "role": "narrator"}}]
|
| 61 |
"""
|
| 62 |
|
|
@@ -69,15 +63,13 @@ def analyze_text_with_tiny_ai(text):
|
|
| 69 |
outputs = pipe(messages)
|
| 70 |
result_text = outputs[0]["generated_text"][-1]["content"]
|
| 71 |
|
| 72 |
-
#
|
| 73 |
json_match = re.search(r'\[.*\]', result_text, re.DOTALL)
|
| 74 |
if json_match:
|
| 75 |
json_str = json_match.group(0)
|
| 76 |
-
|
| 77 |
-
return data
|
| 78 |
else:
|
| 79 |
-
|
| 80 |
-
print(f"⚠️ Модель ответила не JSON: {result_text}")
|
| 81 |
return [{"text": text, "role": "narrator"}]
|
| 82 |
|
| 83 |
except Exception as e:
|
|
@@ -90,44 +82,46 @@ async def generate_segment(text, role):
|
|
| 90 |
if not text.strip(): return None
|
| 91 |
conf = VOICE_CONFIG.get(role, VOICE_CONFIG["narrator"])
|
| 92 |
path = os.path.join(TEMP_DIR, f"{uuid.uuid4().hex}.mp3")
|
|
|
|
| 93 |
try:
|
| 94 |
comm = edge_tts.Communicate(text, conf["voice"], rate=conf["rate"], pitch=conf["pitch"])
|
| 95 |
await comm.save(path)
|
| 96 |
return path
|
| 97 |
-
except:
|
|
|
|
| 98 |
|
| 99 |
async def process_book(text):
|
| 100 |
-
if not text.strip(): raise gr.Warning("
|
| 101 |
|
| 102 |
-
print("⚡
|
| 103 |
segments = analyze_text_with_tiny_ai(text)
|
| 104 |
-
print(f"Результат анализа: {len(segments)} кусков.")
|
| 105 |
|
| 106 |
full_audio = AudioSegment.empty()
|
| 107 |
temp_files = []
|
| 108 |
|
| 109 |
progress = gr.Progress()
|
| 110 |
for item in progress.tqdm(segments, desc="Озвучка"):
|
| 111 |
-
#
|
| 112 |
-
if isinstance(item,
|
| 113 |
-
txt, role = item, "narrator"
|
| 114 |
-
else:
|
| 115 |
txt = item.get("text", "")
|
| 116 |
role = item.get("role", "narrator")
|
|
|
|
|
|
|
|
|
|
| 117 |
|
| 118 |
path = await generate_segment(txt, role)
|
| 119 |
|
| 120 |
if path:
|
| 121 |
temp_files.append(path)
|
| 122 |
seg = AudioSegment.from_mp3(path)
|
| 123 |
-
#
|
| 124 |
if len(full_audio) > 0:
|
| 125 |
full_audio = full_audio.append(seg, crossfade=50)
|
| 126 |
else:
|
| 127 |
full_audio = seg
|
| 128 |
await asyncio.sleep(0.1)
|
| 129 |
|
| 130 |
-
out_path = os.path.join(TEMP_DIR, f"
|
| 131 |
full_audio.export(out_path, format="mp3")
|
| 132 |
|
| 133 |
for f in temp_files:
|
|
@@ -137,16 +131,27 @@ async def process_book(text):
|
|
| 137 |
return out_path, segments
|
| 138 |
|
| 139 |
# --- ИНТЕРФЕЙС ---
|
| 140 |
-
css = "body {background-color: #1e1e2e; color: #cdd6f4;} .gradio-container {font-family: 'Verdana', sans-serif;}"
|
| 141 |
-
theme = gr.themes.Soft(primary_hue="indigo")
|
| 142 |
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
|
|
|
|
|
|
| 150 |
|
| 151 |
with gr.Row():
|
| 152 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
from pydub import AudioSegment
|
| 10 |
from transformers import pipeline
|
| 11 |
|
| 12 |
+
# --- НАСТРОЙКИ ГОЛОСОВ (ФЭНТЕЗИ) ---
|
| 13 |
VOICE_CONFIG = {
|
| 14 |
+
"narrator": {"voice": "ru-RU-DmitryNeural", "pitch": "-7Hz", "rate": "-5%"},
|
| 15 |
+
"male": {"voice": "ru-RU-DenisNeural", "pitch": "-2Hz", "rate": "+0%"},
|
| 16 |
+
"female": {"voice": "ru-RU-SvetlanaNeural","pitch": "+5Hz", "rate": "+5%"}
|
| 17 |
}
|
| 18 |
|
| 19 |
TEMP_DIR = tempfile.gettempdir()
|
| 20 |
|
| 21 |
# --- ЗАГРУЗКА МАЛЕНЬКОЙ НЕЙРОСЕТИ ---
|
| 22 |
+
# Qwen 2.5 0.5B Instruct - очень легкая, но умная
|
| 23 |
MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct"
|
| 24 |
|
| 25 |
+
print(f"🚀 Загрузка модели {MODEL_ID}...")
|
| 26 |
try:
|
|
|
|
| 27 |
pipe = pipeline(
|
| 28 |
"text-generation",
|
| 29 |
model=MODEL_ID,
|
| 30 |
+
device_map="auto",
|
| 31 |
max_new_tokens=2048,
|
| 32 |
trust_remote_code=True
|
| 33 |
)
|
| 34 |
+
print("✅ Модель готова!")
|
| 35 |
except Exception as e:
|
| 36 |
print(f"❌ Ошибка загрузки модели: {e}")
|
| 37 |
pipe = None
|
| 38 |
|
| 39 |
def analyze_text_with_tiny_ai(text):
|
| 40 |
+
"""Анализ текста легкой нейросетью."""
|
|
|
|
|
|
|
| 41 |
if not pipe:
|
| 42 |
return [{"text": text, "role": "narrator"}]
|
| 43 |
|
|
|
|
|
|
|
| 44 |
system_prompt = (
|
| 45 |
+
"Ты редактор. Твоя задача - определить роль для озвучки.\n"
|
| 46 |
+
"Роли: narrator (автор), male (мужчина), female (женщина).\n"
|
| 47 |
+
"Верни ТОЛЬКО JSON список."
|
| 48 |
)
|
| 49 |
|
| 50 |
+
user_prompt = f"""Разбей текст на роли:
|
|
|
|
| 51 |
"{text}"
|
| 52 |
|
| 53 |
+
Пример JSON ответа:
|
| 54 |
[{{"text": "- Привет", "role": "male"}}, {{"text": "- сказала она", "role": "narrator"}}]
|
| 55 |
"""
|
| 56 |
|
|
|
|
| 63 |
outputs = pipe(messages)
|
| 64 |
result_text = outputs[0]["generated_text"][-1]["content"]
|
| 65 |
|
| 66 |
+
# Поиск JSON в ответе
|
| 67 |
json_match = re.search(r'\[.*\]', result_text, re.DOTALL)
|
| 68 |
if json_match:
|
| 69 |
json_str = json_match.group(0)
|
| 70 |
+
return json.loads(json_str)
|
|
|
|
| 71 |
else:
|
| 72 |
+
print(f"⚠️ Не JSON: {result_text}")
|
|
|
|
| 73 |
return [{"text": text, "role": "narrator"}]
|
| 74 |
|
| 75 |
except Exception as e:
|
|
|
|
| 82 |
if not text.strip(): return None
|
| 83 |
conf = VOICE_CONFIG.get(role, VOICE_CONFIG["narrator"])
|
| 84 |
path = os.path.join(TEMP_DIR, f"{uuid.uuid4().hex}.mp3")
|
| 85 |
+
|
| 86 |
try:
|
| 87 |
comm = edge_tts.Communicate(text, conf["voice"], rate=conf["rate"], pitch=conf["pitch"])
|
| 88 |
await comm.save(path)
|
| 89 |
return path
|
| 90 |
+
except:
|
| 91 |
+
return None
|
| 92 |
|
| 93 |
async def process_book(text):
|
| 94 |
+
if not text.strip(): raise gr.Warning("Введите текст!")
|
| 95 |
|
| 96 |
+
print("⚡ Анализ текста...")
|
| 97 |
segments = analyze_text_with_tiny_ai(text)
|
|
|
|
| 98 |
|
| 99 |
full_audio = AudioSegment.empty()
|
| 100 |
temp_files = []
|
| 101 |
|
| 102 |
progress = gr.Progress()
|
| 103 |
for item in progress.tqdm(segments, desc="Озвучка"):
|
| 104 |
+
# Защита от некорректного формата
|
| 105 |
+
if isinstance(item, dict):
|
|
|
|
|
|
|
| 106 |
txt = item.get("text", "")
|
| 107 |
role = item.get("role", "narrator")
|
| 108 |
+
else:
|
| 109 |
+
txt = str(item)
|
| 110 |
+
role = "narrator"
|
| 111 |
|
| 112 |
path = await generate_segment(txt, role)
|
| 113 |
|
| 114 |
if path:
|
| 115 |
temp_files.append(path)
|
| 116 |
seg = AudioSegment.from_mp3(path)
|
| 117 |
+
# Плавная склейка (50ms)
|
| 118 |
if len(full_audio) > 0:
|
| 119 |
full_audio = full_audio.append(seg, crossfade=50)
|
| 120 |
else:
|
| 121 |
full_audio = seg
|
| 122 |
await asyncio.sleep(0.1)
|
| 123 |
|
| 124 |
+
out_path = os.path.join(TEMP_DIR, f"fantasy_{uuid.uuid4().hex}.mp3")
|
| 125 |
full_audio.export(out_path, format="mp3")
|
| 126 |
|
| 127 |
for f in temp_files:
|
|
|
|
| 131 |
return out_path, segments
|
| 132 |
|
| 133 |
# --- ИНТЕРФЕЙС ---
|
|
|
|
|
|
|
| 134 |
|
| 135 |
+
css = """
|
| 136 |
+
body {background-color: #111827; color: #e5e7eb;}
|
| 137 |
+
.container {max-width: 900px; margin: auto;}
|
| 138 |
+
"""
|
| 139 |
+
|
| 140 |
+
theme = gr.themes.Soft(primary_hue="indigo", secondary_hue="slate")
|
| 141 |
+
|
| 142 |
+
with gr.Blocks(theme=theme, css=css, title="Fantasy Lite TTS") as demo:
|
| 143 |
+
gr.Markdown("# ⚡ Fantasy Lite TTS (Qwen 0.5B)")
|
| 144 |
|
| 145 |
with gr.Row():
|
| 146 |
+
with gr.Column(scale=2):
|
| 147 |
+
inp = gr.Textbox(label="Текст", lines=10, placeholder="Вставьте текст...", value='— Кто здесь? — спросил рыцарь.\nВедьма усмехнулась: — Твоя судьба.')
|
| 148 |
+
btn = gr.Button("🚀 Создать", variant="primary")
|
| 149 |
+
|
| 150 |
+
with gr.Column(scale=1):
|
| 151 |
+
out_audio = gr.Audio(label="Результат", type="filepath")
|
| 152 |
+
out_debug = gr.JSON(label="Лог нейросети")
|
| 153 |
+
|
| 154 |
+
btn.click(process_book, inputs=inp, outputs=[out_audio, out_debug])
|
| 155 |
+
|
| 156 |
+
if __name__ == "__main__":
|
| 157 |
+
demo.queue().launch()
|