Update app.py
Browse files
app.py
CHANGED
|
@@ -32,6 +32,9 @@ ALLOWED_EXTENSIONS = ALLOWED_AUDIO_EXTENSIONS.union(ALLOWED_VIDEO_EXTENSIONS)
|
|
| 32 |
API_KEY = os.environ.get("API_KEY") # 在 HF Space 的 Repo secrets 設定
|
| 33 |
MODEL_NAME = os.environ.get("WHISPER_MODEL", "guillaumekln/faster-whisper-large-v2")
|
| 34 |
|
|
|
|
|
|
|
|
|
|
| 35 |
# ------------------------------------
|
| 36 |
# 裝置與模型
|
| 37 |
# ------------------------------------
|
|
@@ -64,7 +67,7 @@ active_requests = 0
|
|
| 64 |
# ------------------------------------
|
| 65 |
def validate_api_key(req):
|
| 66 |
api_key = req.headers.get('X-API-Key')
|
| 67 |
-
return api_key == API_KEY if API_KEY else
|
| 68 |
|
| 69 |
def allowed_file(filename):
|
| 70 |
return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS
|
|
@@ -83,7 +86,7 @@ def extract_audio_from_video(video_path, output_audio_path):
|
|
| 83 |
使用 ffmpeg 從影片擷取 PCM WAV,並用 moviepy 檢查長度
|
| 84 |
"""
|
| 85 |
try:
|
| 86 |
-
#
|
| 87 |
ffmpeg.input(video_path).output(
|
| 88 |
output_audio_path,
|
| 89 |
acodec='pcm_s16le'
|
|
@@ -94,17 +97,17 @@ def extract_audio_from_video(video_path, output_audio_path):
|
|
| 94 |
video = VideoFileClip(video_path)
|
| 95 |
if video.duration > MAX_FILE_DURATION:
|
| 96 |
video.close()
|
| 97 |
-
raise ValueError(f"
|
| 98 |
video.close()
|
| 99 |
|
| 100 |
return output_audio_path
|
| 101 |
except Exception as e:
|
| 102 |
-
logging.exception("
|
| 103 |
-
raise Exception(f"
|
| 104 |
|
| 105 |
def fmt_mmss_mmm(seconds: float) -> str:
|
| 106 |
"""
|
| 107 |
-
轉成 MM:SS.mmm
|
| 108 |
若未來需要小時欄位,可改為 HH:MM:SS.mmm。
|
| 109 |
"""
|
| 110 |
if seconds is None:
|
|
@@ -114,17 +117,26 @@ def fmt_mmss_mmm(seconds: float) -> str:
|
|
| 114 |
sec, ms = divmod(ms, 1000)
|
| 115 |
return f"{minutes:02d}:{sec:02d}.{ms:03d}"
|
| 116 |
|
| 117 |
-
def
|
| 118 |
"""
|
| 119 |
-
讀取 ?lang=
|
| 120 |
"""
|
| 121 |
lang_param = request.args.get("lang", "").strip()
|
| 122 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 123 |
|
| 124 |
def run_transcribe_pipeline(uploaded_file_path: str, file_extension: str):
|
| 125 |
"""
|
| 126 |
-
|
| 127 |
-
回傳:(segments_iterable, is_video)
|
| 128 |
"""
|
| 129 |
is_video = file_extension in ALLOWED_VIDEO_EXTENSIONS
|
| 130 |
temp_audio_path = None
|
|
@@ -135,36 +147,40 @@ def run_transcribe_pipeline(uploaded_file_path: str, file_extension: str):
|
|
| 135 |
transcription_file = temp_audio_path
|
| 136 |
else:
|
| 137 |
transcription_file = uploaded_file_path
|
| 138 |
-
#
|
| 139 |
try:
|
| 140 |
waveform, sample_rate = torchaudio.load(transcription_file, format=file_extension)
|
| 141 |
duration = waveform.size(1) / sample_rate
|
| 142 |
if duration > MAX_FILE_DURATION:
|
| 143 |
-
raise ValueError(f"
|
| 144 |
except Exception:
|
| 145 |
-
logging.exception(f"使用 torchaudio.load
|
| 146 |
try:
|
| 147 |
torchaudio.set_audio_backend("soundfile")
|
| 148 |
waveform, sample_rate = torchaudio.load(transcription_file)
|
| 149 |
duration = waveform.size(1) / sample_rate
|
| 150 |
if duration > MAX_FILE_DURATION:
|
| 151 |
-
raise ValueError(f"
|
| 152 |
except Exception as soundfile_err:
|
| 153 |
-
logging.exception(f"使用 soundfile
|
| 154 |
-
raise Exception(f'
|
| 155 |
finally:
|
| 156 |
torchaudio.set_audio_backend("default")
|
| 157 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 158 |
# 轉錄(保留 segment 級時間)
|
| 159 |
-
language = read_lang_param()
|
| 160 |
segments, info = wmodel.transcribe(
|
| 161 |
transcription_file,
|
| 162 |
beam_size=beamsize,
|
| 163 |
vad_filter=True,
|
| 164 |
-
without_timestamps=False,
|
| 165 |
compression_ratio_threshold=2.4,
|
| 166 |
-
word_timestamps=False,
|
| 167 |
-
language=language
|
|
|
|
| 168 |
)
|
| 169 |
|
| 170 |
return segments, is_video, temp_audio_path
|
|
@@ -182,7 +198,9 @@ def health_check():
|
|
| 182 |
'active_requests': active_requests,
|
| 183 |
'max_duration_supported': MAX_FILE_DURATION,
|
| 184 |
'supported_formats': list(ALLOWED_EXTENSIONS),
|
| 185 |
-
'model': MODEL_NAME
|
|
|
|
|
|
|
| 186 |
})
|
| 187 |
|
| 188 |
@app.route("/status/busy", methods=["GET"])
|
|
@@ -208,7 +226,7 @@ def transcribe_json():
|
|
| 208 |
return jsonify({'error': '伺服器繁忙'}), 503
|
| 209 |
|
| 210 |
active_requests += 1
|
| 211 |
-
|
| 212 |
temp_file_path = None
|
| 213 |
temp_audio_path = None
|
| 214 |
|
|
@@ -260,7 +278,7 @@ def transcribe_json():
|
|
| 260 |
cleanup_temp_files(temp_file_path, temp_audio_path)
|
| 261 |
active_requests -= 1
|
| 262 |
request_semaphore.release()
|
| 263 |
-
logging.info(f"
|
| 264 |
|
| 265 |
# ------------------------------------
|
| 266 |
# 端點 2:純文字(整段合併,沒有時間戳)
|
|
@@ -276,7 +294,7 @@ def transcribe_text_only():
|
|
| 276 |
return jsonify({'error': '伺服器繁忙'}), 503
|
| 277 |
|
| 278 |
active_requests += 1
|
| 279 |
-
|
| 280 |
temp_file_path = None
|
| 281 |
temp_audio_path = None
|
| 282 |
|
|
@@ -317,7 +335,7 @@ def transcribe_text_only():
|
|
| 317 |
cleanup_temp_files(temp_file_path, temp_audio_path)
|
| 318 |
active_requests -= 1
|
| 319 |
request_semaphore.release()
|
| 320 |
-
logging.info(f"
|
| 321 |
|
| 322 |
|
| 323 |
if __name__ == "__main__":
|
|
@@ -325,4 +343,4 @@ if __name__ == "__main__":
|
|
| 325 |
os.makedirs(TEMPORARY_FOLDER)
|
| 326 |
logging.info(f"新建暫存檔案夾: {TEMPORARY_FOLDER}")
|
| 327 |
|
| 328 |
-
app.run(host="0.0.0.0", port=7860, threaded=True)
|
|
|
|
| 32 |
API_KEY = os.environ.get("API_KEY") # 在 HF Space 的 Repo secrets 設定
|
| 33 |
MODEL_NAME = os.environ.get("WHISPER_MODEL", "guillaumekln/faster-whisper-large-v2")
|
| 34 |
|
| 35 |
+
# 預設提示(可用 ?prompt 覆蓋)
|
| 36 |
+
DEFAULT_INITIAL_PROMPT = "請使用繁體中文輸出"
|
| 37 |
+
|
| 38 |
# ------------------------------------
|
| 39 |
# 裝置與模型
|
| 40 |
# ------------------------------------
|
|
|
|
| 67 |
# ------------------------------------
|
| 68 |
def validate_api_key(req):
|
| 69 |
api_key = req.headers.get('X-API-Key')
|
| 70 |
+
return api_key == API_KEY if API_KEY else True # 若沒設定 API_KEY,預設放行(可依需求改)
|
| 71 |
|
| 72 |
def allowed_file(filename):
|
| 73 |
return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS
|
|
|
|
| 86 |
使用 ffmpeg 從影片擷取 PCM WAV,並用 moviepy 檢查長度
|
| 87 |
"""
|
| 88 |
try:
|
| 89 |
+
# 先擷取音訊
|
| 90 |
ffmpeg.input(video_path).output(
|
| 91 |
output_audio_path,
|
| 92 |
acodec='pcm_s16le'
|
|
|
|
| 97 |
video = VideoFileClip(video_path)
|
| 98 |
if video.duration > MAX_FILE_DURATION:
|
| 99 |
video.close()
|
| 100 |
+
raise ValueError(f"視頻時長超過 {MAX_FILE_DURATION} 秒")
|
| 101 |
video.close()
|
| 102 |
|
| 103 |
return output_audio_path
|
| 104 |
except Exception as e:
|
| 105 |
+
logging.exception("提取視頻中的音訊出錯")
|
| 106 |
+
raise Exception(f"提取視頻中的音訊出錯: {str(e)}")
|
| 107 |
|
| 108 |
def fmt_mmss_mmm(seconds: float) -> str:
|
| 109 |
"""
|
| 110 |
+
轉成 MM:SS.mmm(符合需求,如 00:01.000)
|
| 111 |
若未來需要小時欄位,可改為 HH:MM:SS.mmm。
|
| 112 |
"""
|
| 113 |
if seconds is None:
|
|
|
|
| 117 |
sec, ms = divmod(ms, 1000)
|
| 118 |
return f"{minutes:02d}:{sec:02d}.{ms:03d}"
|
| 119 |
|
| 120 |
+
def read_lang_param_with_default_zh():
|
| 121 |
"""
|
| 122 |
+
讀取 ?lang= 參數;沒帶或為 auto 時預設繁體中文 (zh)
|
| 123 |
"""
|
| 124 |
lang_param = request.args.get("lang", "").strip()
|
| 125 |
+
if not lang_param or lang_param.lower() == "auto":
|
| 126 |
+
return "zh"
|
| 127 |
+
return lang_param
|
| 128 |
+
|
| 129 |
+
def read_initial_prompt():
|
| 130 |
+
"""
|
| 131 |
+
讀取 ?prompt= 參數;沒帶則使用 DEFAULT_INITIAL_PROMPT
|
| 132 |
+
"""
|
| 133 |
+
prompt = request.args.get("prompt", "").strip()
|
| 134 |
+
return prompt if prompt else DEFAULT_INITIAL_PROMPT
|
| 135 |
|
| 136 |
def run_transcribe_pipeline(uploaded_file_path: str, file_extension: str):
|
| 137 |
"""
|
| 138 |
+
共用的轉錄流程:處理影片/音訊、長度檢查、呼叫 Faster-Whisper。
|
| 139 |
+
回傳:(segments_iterable, is_video, temp_audio_path)
|
| 140 |
"""
|
| 141 |
is_video = file_extension in ALLOWED_VIDEO_EXTENSIONS
|
| 142 |
temp_audio_path = None
|
|
|
|
| 147 |
transcription_file = temp_audio_path
|
| 148 |
else:
|
| 149 |
transcription_file = uploaded_file_path
|
| 150 |
+
# 檢查音訊長度
|
| 151 |
try:
|
| 152 |
waveform, sample_rate = torchaudio.load(transcription_file, format=file_extension)
|
| 153 |
duration = waveform.size(1) / sample_rate
|
| 154 |
if duration > MAX_FILE_DURATION:
|
| 155 |
+
raise ValueError(f"音訊時長超過 {MAX_FILE_DURATION} 秒")
|
| 156 |
except Exception:
|
| 157 |
+
logging.exception(f"使用 torchaudio.load 載入音訊檔出錯: {transcription_file}")
|
| 158 |
try:
|
| 159 |
torchaudio.set_audio_backend("soundfile")
|
| 160 |
waveform, sample_rate = torchaudio.load(transcription_file)
|
| 161 |
duration = waveform.size(1) / sample_rate
|
| 162 |
if duration > MAX_FILE_DURATION:
|
| 163 |
+
raise ValueError(f"音訊時長超過 {MAX_FILE_DURATION} 秒")
|
| 164 |
except Exception as soundfile_err:
|
| 165 |
+
logging.exception(f"使用 soundfile 後端載入音訊檔出錯: {transcription_file}")
|
| 166 |
+
raise Exception(f'使用兩個後端載入音訊檔都出錯: {str(soundfile_err)}')
|
| 167 |
finally:
|
| 168 |
torchaudio.set_audio_backend("default")
|
| 169 |
|
| 170 |
+
# 預設語言 zh,並帶 initial_prompt(可被 ?lang / ?prompt 覆蓋)
|
| 171 |
+
language = read_lang_param_with_default_zh()
|
| 172 |
+
initial_prompt = read_initial_prompt()
|
| 173 |
+
|
| 174 |
# 轉錄(保留 segment 級時間)
|
|
|
|
| 175 |
segments, info = wmodel.transcribe(
|
| 176 |
transcription_file,
|
| 177 |
beam_size=beamsize,
|
| 178 |
vad_filter=True,
|
| 179 |
+
without_timestamps=False, # 要保留時間戳
|
| 180 |
compression_ratio_threshold=2.4,
|
| 181 |
+
word_timestamps=False, # 如需字級,設 True
|
| 182 |
+
language=language,
|
| 183 |
+
initial_prompt=initial_prompt
|
| 184 |
)
|
| 185 |
|
| 186 |
return segments, is_video, temp_audio_path
|
|
|
|
| 198 |
'active_requests': active_requests,
|
| 199 |
'max_duration_supported': MAX_FILE_DURATION,
|
| 200 |
'supported_formats': list(ALLOWED_EXTENSIONS),
|
| 201 |
+
'model': MODEL_NAME,
|
| 202 |
+
'default_language': 'zh',
|
| 203 |
+
'default_initial_prompt': DEFAULT_INITIAL_PROMPT
|
| 204 |
})
|
| 205 |
|
| 206 |
@app.route("/status/busy", methods=["GET"])
|
|
|
|
| 226 |
return jsonify({'error': '伺服器繁忙'}), 503
|
| 227 |
|
| 228 |
active_requests += 1
|
| 229 |
+
t0 = time.time()
|
| 230 |
temp_file_path = None
|
| 231 |
temp_audio_path = None
|
| 232 |
|
|
|
|
| 278 |
cleanup_temp_files(temp_file_path, temp_audio_path)
|
| 279 |
active_requests -= 1
|
| 280 |
request_semaphore.release()
|
| 281 |
+
logging.info(f"/whisper_transcribe 用時:{time.time() - t0:.2f}s (活動請求:{active_requests})")
|
| 282 |
|
| 283 |
# ------------------------------------
|
| 284 |
# 端點 2:純文字(整段合併,沒有時間戳)
|
|
|
|
| 294 |
return jsonify({'error': '伺服器繁忙'}), 503
|
| 295 |
|
| 296 |
active_requests += 1
|
| 297 |
+
t0 = time.time()
|
| 298 |
temp_file_path = None
|
| 299 |
temp_audio_path = None
|
| 300 |
|
|
|
|
| 335 |
cleanup_temp_files(temp_file_path, temp_audio_path)
|
| 336 |
active_requests -= 1
|
| 337 |
request_semaphore.release()
|
| 338 |
+
logging.info(f"/whisper_transcribe_text 用時:{time.time() - t0:.2f}s (活動請求:{active_requests})")
|
| 339 |
|
| 340 |
|
| 341 |
if __name__ == "__main__":
|
|
|
|
| 343 |
os.makedirs(TEMPORARY_FOLDER)
|
| 344 |
logging.info(f"新建暫存檔案夾: {TEMPORARY_FOLDER}")
|
| 345 |
|
| 346 |
+
app.run(host="0.0.0.0", port=7860, threaded=True)
|