Spaces:
Sleeping
Sleeping
Update core.py
Browse files
core.py
CHANGED
|
@@ -16,18 +16,54 @@ def parse_srt(srt_path):
|
|
| 16 |
print(f"解析字幕檔案錯誤: {e}")
|
| 17 |
return None
|
| 18 |
|
| 19 |
-
#
|
| 20 |
def preprocess_transcript(transcript):
|
| 21 |
-
|
| 22 |
transcript = re.sub(r'<[^>]*?>', '', transcript)
|
|
|
|
| 23 |
# 移除多個連續空格
|
| 24 |
transcript = re.sub(r'\s+', ' ', transcript)
|
| 25 |
-
|
| 26 |
-
|
|
|
|
|
|
|
| 27 |
# 移除文字前後的空白
|
| 28 |
transcript = transcript.strip()
|
| 29 |
return transcript
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
def process_files(api_key, test_transcript_file, test_srt_file, batch_size):
|
| 32 |
# 1. 配置 Gemini API
|
| 33 |
try:
|
|
@@ -48,67 +84,122 @@ def process_files(api_key, test_transcript_file, test_srt_file, batch_size):
|
|
| 48 |
if not os.path.exists(test_srt_file):
|
| 49 |
return f"錯誤: 找不到口語稿檔案: {test_srt_file},請確認您已上傳該檔案", None, None
|
| 50 |
else:
|
| 51 |
-
|
| 52 |
-
if
|
| 53 |
return "錯誤: 解析口語稿檔案失敗,請確認檔案內容", None, None
|
| 54 |
|
|
|
|
|
|
|
|
|
|
| 55 |
# 儲存所有報告
|
| 56 |
all_reports = []
|
| 57 |
keys = list(srt_data.keys()) # 取得編號
|
| 58 |
|
|
|
|
|
|
|
|
|
|
| 59 |
# 處理每個批次
|
| 60 |
-
for i in range(0, len(keys), batch_size):
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
prompt = (
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
try:
|
| 81 |
response = model.generate_content(prompt)
|
| 82 |
corrected_subtitle = response.text
|
| 83 |
-
print(f"第 {i // batch_size + 1} 批次 Gemini 模型的回應:")
|
| 84 |
print(corrected_subtitle)
|
| 85 |
|
| 86 |
# 使用 re.split 分割字幕和報告
|
| 87 |
parts = re.split(r'<<<分隔符號>>>', corrected_subtitle, maxsplit=1) # 只分割一次
|
| 88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
report = ""
|
| 90 |
if len(parts) > 1:
|
| 91 |
report = parts[1] # 取得修改清單文字
|
| 92 |
report_lines = report.strip().split('\n')
|
| 93 |
all_reports.extend(report_lines)
|
| 94 |
|
| 95 |
-
for line in corrected_lines:
|
| 96 |
-
match = re.match(r'編號(\d+):(.*)', line)
|
| 97 |
-
if match:
|
| 98 |
-
index = int(match.group(1))
|
| 99 |
-
corrected_text = match.group(2).strip()
|
| 100 |
-
srt_data[index]['text'] = corrected_text
|
| 101 |
except Exception as e:
|
| 102 |
-
return f"第 {i // batch_size + 1} 批次修正過程失敗:{e}", None, None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
# 寫回 SRT 檔案
|
| 105 |
new_subs = []
|
| 106 |
-
for index, data in srt_data.items():
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
|
| 113 |
# 將 new_subs 轉換為 SubRipFile 物件
|
| 114 |
new_srt = pysrt.SubRipFile(items=new_subs)
|
|
|
|
| 16 |
print(f"解析字幕檔案錯誤: {e}")
|
| 17 |
return None
|
| 18 |
|
| 19 |
+
# 改進的預處理逐字稿函數
|
| 20 |
def preprocess_transcript(transcript):
|
| 21 |
+
# 移除 SSML 標籤
|
| 22 |
transcript = re.sub(r'<[^>]*?>', '', transcript)
|
| 23 |
+
|
| 24 |
# 移除多個連續空格
|
| 25 |
transcript = re.sub(r'\s+', ' ', transcript)
|
| 26 |
+
|
| 27 |
+
# 將連續多個換行替換為單個分隔符(而不是完全移除)
|
| 28 |
+
transcript = re.sub(r'\n+', ' | ', transcript)
|
| 29 |
+
|
| 30 |
# 移除文字前後的空白
|
| 31 |
transcript = transcript.strip()
|
| 32 |
return transcript
|
| 33 |
|
| 34 |
+
# 驗證SRT結構
|
| 35 |
+
def validate_srt(original_srt, modified_srt):
|
| 36 |
+
"""
|
| 37 |
+
驗證修改後的SRT是否保留了原始SRT的基本結構
|
| 38 |
+
|
| 39 |
+
Args:
|
| 40 |
+
original_srt: 原始SRT數據
|
| 41 |
+
modified_srt: 修改後的SRT數據
|
| 42 |
+
|
| 43 |
+
Returns:
|
| 44 |
+
(bool, str): (是否有效, 錯誤信息)
|
| 45 |
+
"""
|
| 46 |
+
# 檢查編號完整性
|
| 47 |
+
if set(original_srt.keys()) != set(modified_srt.keys()):
|
| 48 |
+
missing = set(original_srt.keys()) - set(modified_srt.keys())
|
| 49 |
+
extra = set(modified_srt.keys()) - set(original_srt.keys())
|
| 50 |
+
return False, f"編號不匹配: 缺少 {missing}, 多出 {extra}"
|
| 51 |
+
|
| 52 |
+
# 檢查時間戳保持不變
|
| 53 |
+
for idx in original_srt:
|
| 54 |
+
if original_srt[idx]['time'] != modified_srt[idx]['time']:
|
| 55 |
+
return False, f"編號 {idx} 的時間戳被修改"
|
| 56 |
+
|
| 57 |
+
# 檢查文字長度變化不太大(防止大規模刪減或添加)
|
| 58 |
+
for idx in original_srt:
|
| 59 |
+
orig_len = len(original_srt[idx]['text'])
|
| 60 |
+
mod_len = len(modified_srt[idx]['text'])
|
| 61 |
+
# 允許30%的文字長度變化
|
| 62 |
+
if abs(orig_len - mod_len) / max(1, orig_len) > 0.3:
|
| 63 |
+
return False, f"編號 {idx} 的文字長度變化過大: 原 {orig_len}, 現 {mod_len}"
|
| 64 |
+
|
| 65 |
+
return True, "驗證通過"
|
| 66 |
+
|
| 67 |
def process_files(api_key, test_transcript_file, test_srt_file, batch_size):
|
| 68 |
# 1. 配置 Gemini API
|
| 69 |
try:
|
|
|
|
| 84 |
if not os.path.exists(test_srt_file):
|
| 85 |
return f"錯誤: 找不到口語稿檔案: {test_srt_file},請確認您已上傳該檔案", None, None
|
| 86 |
else:
|
| 87 |
+
original_srt_data = parse_srt(test_srt_file)
|
| 88 |
+
if original_srt_data is None:
|
| 89 |
return "錯誤: 解析口語稿檔案失敗,請確認檔案內容", None, None
|
| 90 |
|
| 91 |
+
# 創建工作副本
|
| 92 |
+
srt_data = original_srt_data.copy()
|
| 93 |
+
|
| 94 |
# 儲存所有報告
|
| 95 |
all_reports = []
|
| 96 |
keys = list(srt_data.keys()) # 取得編號
|
| 97 |
|
| 98 |
+
# 使用重疊批次而非固定批次
|
| 99 |
+
overlap = max(2, batch_size // 4) # 25% 的重疊
|
| 100 |
+
|
| 101 |
# 處理每個批次
|
| 102 |
+
for i in range(0, len(keys), batch_size - overlap):
|
| 103 |
+
end_idx = min(i + batch_size, len(keys))
|
| 104 |
+
batch_keys = keys[i:end_idx]
|
| 105 |
+
|
| 106 |
+
# 確保前一批次的最後部分有重疊
|
| 107 |
+
context_start = max(0, i - overlap)
|
| 108 |
+
context_keys = keys[context_start:i] if i > 0 else []
|
| 109 |
+
|
| 110 |
+
# 準備本批次處理的數據
|
| 111 |
+
batch_with_context = context_keys + batch_keys
|
| 112 |
+
context_batch_data = {key: srt_data[key] for key in batch_with_context}
|
| 113 |
+
|
| 114 |
+
# 在提示中標記哪些是實際需要處理的部分(非上下文)
|
| 115 |
+
subtitle_lines = []
|
| 116 |
+
for key in batch_with_context:
|
| 117 |
+
prefix = "處理→ " if key in batch_keys else "上下文: "
|
| 118 |
+
subtitle_lines.append(f"{prefix}編號{key}:{context_batch_data[key]['text']}")
|
| 119 |
+
|
| 120 |
+
subtitle_content = "\n".join(subtitle_lines)
|
| 121 |
+
|
| 122 |
prompt = (
|
| 123 |
+
"請扮演一位專業的文字編輯,比對以下字幕檔和逐字稿,並以逐字稿的內容為準,修正字幕檔中因語音轉文字造成的錯別字、漏字、多字、詞語誤用等問題,請勿修改語氣和修辭。\n"
|
| 124 |
+
|
| 125 |
+
"【重要規則】:\n"
|
| 126 |
+
"1. 只修正標記有「處理→」的字幕行,標記為「上下文:」的行僅供參考\n"
|
| 127 |
+
"2. 必須嚴格保持原始字幕的時間戳和斷句結構\n"
|
| 128 |
+
"3. 不允許合併任何字幕行,每個編號必須獨立輸出\n"
|
| 129 |
+
"4. 不允許拆分任何字幕行\n"
|
| 130 |
+
"5. 不要新增或移除任何字幕段落\n"
|
| 131 |
+
"6. 即使對於很長的字幕行,也必須完整處理,不能忽略任何部分\n"
|
| 132 |
+
"7. 逐字稿中的 '|' 符號代表可能的段落分隔,請參考但不必完全依照\n"
|
| 133 |
+
|
| 134 |
+
"請使用以下格式輸出每個字幕行:\n"
|
| 135 |
+
"編號X:修正後的文字\n"
|
| 136 |
+
|
| 137 |
+
"接著,請將**當前批次**所有修改的內容整理成一個修改清單,並放在所有字幕內容之後。\n"
|
| 138 |
+
"請使用以下格式輸出修改清單: \n"
|
| 139 |
+
"編號: 原文: 修正後\n"
|
| 140 |
+
"請使用 `<<<分隔符號>>>` 來分隔字幕與修改清單。\n"
|
| 141 |
+
|
| 142 |
+
"字幕檔內容:\n"
|
| 143 |
+
f"{subtitle_content}\n\n"
|
| 144 |
+
"逐字稿內容:\n"
|
| 145 |
+
f"{transcript_content}\n\n"
|
| 146 |
+
)
|
| 147 |
+
|
| 148 |
try:
|
| 149 |
response = model.generate_content(prompt)
|
| 150 |
corrected_subtitle = response.text
|
| 151 |
+
print(f"第 {i // (batch_size - overlap) + 1} 批次 Gemini 模型的回應:")
|
| 152 |
print(corrected_subtitle)
|
| 153 |
|
| 154 |
# 使用 re.split 分割字幕和報告
|
| 155 |
parts = re.split(r'<<<分隔符號>>>', corrected_subtitle, maxsplit=1) # 只分割一次
|
| 156 |
+
|
| 157 |
+
# 改進的字幕解析方法
|
| 158 |
+
corrected_lines = parts[0].strip().split('\n')
|
| 159 |
+
|
| 160 |
+
# 建立一個映射表來追踪已處理的編號
|
| 161 |
+
processed_indices = set()
|
| 162 |
+
|
| 163 |
+
for line in corrected_lines:
|
| 164 |
+
# 確保只處理「處理→」標記的編號字幕行
|
| 165 |
+
match = re.match(r'(?:處理→ )?編號(\d+):(.*)', line)
|
| 166 |
+
if match:
|
| 167 |
+
index = int(match.group(1))
|
| 168 |
+
corrected_text = match.group(2).strip()
|
| 169 |
+
|
| 170 |
+
# 確保此編號在當前批次中且需要處理
|
| 171 |
+
if index in batch_keys:
|
| 172 |
+
srt_data[index]['text'] = corrected_text
|
| 173 |
+
processed_indices.add(index)
|
| 174 |
+
|
| 175 |
+
# 檢查是否所有批次中的編號都被處理了
|
| 176 |
+
for index in batch_keys:
|
| 177 |
+
if index not in processed_indices:
|
| 178 |
+
print(f"警告:編號 {index} 在AI處理後丟失,保持原始字幕內容")
|
| 179 |
+
|
| 180 |
+
# 處理報告部分
|
| 181 |
report = ""
|
| 182 |
if len(parts) > 1:
|
| 183 |
report = parts[1] # 取得修改清單文字
|
| 184 |
report_lines = report.strip().split('\n')
|
| 185 |
all_reports.extend(report_lines)
|
| 186 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 187 |
except Exception as e:
|
| 188 |
+
return f"第 {i // (batch_size - overlap) + 1} 批次修正過程失敗:{e}", None, None
|
| 189 |
+
|
| 190 |
+
# 驗證修改後的SRT結構
|
| 191 |
+
is_valid, error_msg = validate_srt(original_srt_data, srt_data)
|
| 192 |
+
if not is_valid:
|
| 193 |
+
return f"SRT結構驗證失敗: {error_msg}", None, None
|
| 194 |
|
| 195 |
# 寫回 SRT 檔案
|
| 196 |
new_subs = []
|
| 197 |
+
for index, data in sorted(srt_data.items()): # 確保按編號順序排序
|
| 198 |
+
time_str = data['time']
|
| 199 |
+
start_str, end_str = time_str.split(" --> ")
|
| 200 |
+
start = pysrt.srtitem.SubRipTime.from_string(start_str)
|
| 201 |
+
end = pysrt.srtitem.SubRipTime.from_string(end_str)
|
| 202 |
+
new_subs.append(pysrt.SubRipItem(index=index, start=start, end=end, text=data['text']))
|
| 203 |
|
| 204 |
# 將 new_subs 轉換為 SubRipFile 物件
|
| 205 |
new_srt = pysrt.SubRipFile(items=new_subs)
|