Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -17,7 +17,7 @@ from fastapi import FastAPI, HTTPException, UploadFile, File, Form, BackgroundTa
|
|
17 |
from fastapi.responses import JSONResponse
|
18 |
from fastapi.staticfiles import StaticFiles
|
19 |
from typing import Dict, Any, Optional, Tuple, List
|
20 |
-
from
|
21 |
|
22 |
# Configure logging
|
23 |
logging.basicConfig(level=logging.INFO)
|
@@ -35,7 +35,8 @@ models_loaded = False
|
|
35 |
loading_in_progress = False
|
36 |
loading_thread = None
|
37 |
model_status = {
|
38 |
-
"
|
|
|
39 |
"mt": "not_loaded",
|
40 |
"tts": "not_loaded"
|
41 |
}
|
@@ -43,10 +44,10 @@ error_message = None
|
|
43 |
current_tts_language = "tgl" # Track the current TTS language
|
44 |
|
45 |
# Model instances
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
mt_model = None
|
51 |
mt_tokenizer = None
|
52 |
tts_model = None
|
@@ -62,11 +63,9 @@ LANGUAGE_MAPPING = {
|
|
62 |
"Pangasinan": "pag"
|
63 |
}
|
64 |
|
65 |
-
#
|
66 |
-
|
67 |
-
|
68 |
-
"tgl": "tagalog"
|
69 |
-
}
|
70 |
|
71 |
NLLB_LANGUAGE_CODES = {
|
72 |
"eng": "eng_Latn",
|
@@ -93,39 +92,60 @@ def check_inappropriate_content(text: str) -> bool:
|
|
93 |
Check if the text contains inappropriate content.
|
94 |
Returns True if inappropriate content is detected, False otherwise.
|
95 |
"""
|
|
|
96 |
text_lower = text.lower()
|
|
|
|
|
97 |
for word in INAPPROPRIATE_WORDS:
|
|
|
98 |
pattern = r'\b' + re.escape(word) + r'\b'
|
99 |
if re.search(pattern, text_lower):
|
100 |
logger.warning(f"Inappropriate content detected: {word}")
|
101 |
return True
|
|
|
102 |
return False
|
103 |
|
104 |
# Function to save PCM data as a WAV file
|
105 |
def save_pcm_to_wav(pcm_data: list, sample_rate: int, output_path: str):
|
|
|
106 |
pcm_array = np.array(pcm_data, dtype=np.int16)
|
|
|
107 |
with wave.open(output_path, 'wb') as wav_file:
|
|
|
108 |
wav_file.setnchannels(1)
|
109 |
-
wav_file.setsampwidth(2)
|
110 |
wav_file.setframerate(sample_rate)
|
|
|
111 |
wav_file.writeframes(pcm_array.tobytes())
|
112 |
|
113 |
# Function to detect speech using an energy-based approach
|
114 |
def detect_speech(waveform: torch.Tensor, sample_rate: int, threshold: float = 0.01, min_speech_duration: float = 0.5) -> bool:
|
|
|
|
|
|
|
|
|
|
|
115 |
waveform_np = waveform.numpy()
|
116 |
if waveform_np.ndim > 1:
|
117 |
-
waveform_np = waveform_np.mean(axis=0)
|
|
|
|
|
118 |
rms = np.sqrt(np.mean(waveform_np**2))
|
119 |
logger.info(f"RMS energy: {rms}")
|
|
|
|
|
120 |
if rms < threshold:
|
121 |
logger.info("No speech detected: RMS energy below threshold")
|
122 |
return False
|
|
|
|
|
|
|
123 |
return True
|
124 |
|
125 |
# Function to clean up old audio files
|
126 |
def cleanup_old_audio_files():
|
127 |
logger.info("Starting cleanup of old audio files...")
|
128 |
-
expiration_time = datetime.now() - timedelta(minutes=10)
|
129 |
for filename in os.listdir(AUDIO_DIR):
|
130 |
file_path = os.path.join(AUDIO_DIR, filename)
|
131 |
if os.path.isfile(file_path):
|
@@ -141,49 +161,53 @@ def cleanup_old_audio_files():
|
|
141 |
def schedule_cleanup():
|
142 |
while True:
|
143 |
cleanup_old_audio_files()
|
144 |
-
time.sleep(300)
|
145 |
|
146 |
# Function to load models in background
|
147 |
def load_models_task():
|
148 |
global models_loaded, loading_in_progress, model_status, error_message
|
149 |
-
global
|
150 |
global mt_model, mt_tokenizer, tts_model, tts_tokenizer
|
151 |
|
152 |
try:
|
153 |
loading_in_progress = True
|
|
|
154 |
|
155 |
-
# Load STT
|
156 |
-
logger.info("Starting to load STT
|
157 |
-
from transformers import
|
158 |
|
159 |
try:
|
160 |
logger.info("Loading Whisper STT model...")
|
161 |
-
model_status["
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
stt_model_whisper.to(device)
|
166 |
logger.info("Whisper STT model loaded successfully")
|
167 |
-
model_status["
|
168 |
-
except Exception as
|
169 |
-
logger.error(f"Failed to load Whisper STT model: {str(
|
170 |
-
model_status["
|
171 |
-
error_message = f"Whisper STT model loading failed: {str(
|
172 |
return
|
173 |
-
|
|
|
|
|
|
|
|
|
174 |
try:
|
175 |
logger.info("Loading MMS STT model...")
|
176 |
-
|
177 |
-
|
178 |
-
|
|
|
179 |
logger.info("MMS STT model loaded successfully")
|
180 |
-
model_status["
|
181 |
-
except Exception as
|
182 |
-
logger.error(f"Failed to load MMS STT model: {str(
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
return
|
187 |
|
188 |
# Load MT model
|
189 |
logger.info("Starting to load MT model...")
|
@@ -203,7 +227,7 @@ def load_models_task():
|
|
203 |
error_message = f"MT model loading failed: {str(e)}"
|
204 |
return
|
205 |
|
206 |
-
# Load TTS model (default to Tagalog)
|
207 |
logger.info("Starting to load TTS model...")
|
208 |
from transformers import VitsModel, AutoTokenizer
|
209 |
|
@@ -217,6 +241,7 @@ def load_models_task():
|
|
217 |
model_status["tts"] = "loaded"
|
218 |
except Exception as e:
|
219 |
logger.error(f"Failed to load TTS model for Tagalog: {str(e)}")
|
|
|
220 |
try:
|
221 |
logger.info("Falling back to MMS-TTS English model...")
|
222 |
tts_model = VitsModel.from_pretrained("facebook/mms-tts-eng")
|
@@ -257,13 +282,21 @@ def start_cleanup_task():
|
|
257 |
|
258 |
# Function to load or update TTS model for a specific language
|
259 |
def load_tts_model_for_language(target_code: str) -> bool:
|
|
|
|
|
|
|
|
|
260 |
global tts_model, tts_tokenizer, current_tts_language, model_status
|
|
|
261 |
if target_code not in LANGUAGE_MAPPING.values():
|
262 |
logger.error(f"Invalid language code: {target_code}")
|
263 |
return False
|
|
|
|
|
264 |
if current_tts_language == target_code and model_status["tts"].startswith("loaded"):
|
265 |
logger.info(f"TTS model for {target_code} is already loaded.")
|
266 |
return True
|
|
|
267 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
268 |
try:
|
269 |
logger.info(f"Loading MMS-TTS model for {target_code}...")
|
@@ -293,21 +326,32 @@ def load_tts_model_for_language(target_code: str) -> bool:
|
|
293 |
|
294 |
# Function to synthesize speech from text
|
295 |
def synthesize_speech(text: str, target_code: str) -> Tuple[Optional[str], Optional[str]]:
|
|
|
|
|
|
|
|
|
296 |
global tts_model, tts_tokenizer
|
|
|
297 |
request_id = str(uuid.uuid4())
|
298 |
output_path = os.path.join(AUDIO_DIR, f"{request_id}.wav")
|
|
|
|
|
299 |
if not load_tts_model_for_language(target_code):
|
300 |
return None, "Failed to load TTS model for the target language"
|
|
|
301 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
302 |
try:
|
303 |
-
inputs = tts_tokenizer(text, return_tensors="pt").
|
304 |
with torch.no_grad():
|
305 |
output = tts_model(**inputs)
|
306 |
speech = output.waveform.cpu().numpy().squeeze()
|
307 |
speech = (speech * 32767).astype(np.int16)
|
308 |
sample_rate = tts_model.config.sampling_rate
|
|
|
|
|
309 |
save_pcm_to_wav(speech.tolist(), sample_rate, output_path)
|
310 |
logger.info(f"Saved synthesized audio to {output_path}")
|
|
|
311 |
return output_path, None
|
312 |
except Exception as e:
|
313 |
error_msg = f"Error during TTS conversion: {str(e)}"
|
@@ -323,11 +367,14 @@ async def startup_event():
|
|
323 |
|
324 |
@app.get("/")
|
325 |
async def root():
|
|
|
326 |
logger.info("Root endpoint requested")
|
327 |
return {"status": "healthy"}
|
328 |
|
329 |
@app.get("/health")
|
330 |
async def health_check():
|
|
|
|
|
331 |
logger.info("Health check requested")
|
332 |
return {
|
333 |
"status": "healthy",
|
@@ -339,16 +386,22 @@ async def health_check():
|
|
339 |
|
340 |
@app.post("/translate-text")
|
341 |
async def translate_text(text: str = Form(...), source_lang: str = Form(...), target_lang: str = Form(...)):
|
|
|
342 |
global mt_model, mt_tokenizer, tts_model, tts_tokenizer, current_tts_language
|
|
|
343 |
if not text:
|
344 |
raise HTTPException(status_code=400, detail="No text provided")
|
345 |
if source_lang not in LANGUAGE_MAPPING or target_lang not in LANGUAGE_MAPPING:
|
346 |
raise HTTPException(status_code=400, detail="Invalid language selected")
|
|
|
347 |
logger.info(f"Translate-text requested: {text} from {source_lang} to {target_lang}")
|
348 |
request_id = str(uuid.uuid4())
|
|
|
|
|
349 |
source_code = LANGUAGE_MAPPING[source_lang]
|
350 |
target_code = LANGUAGE_MAPPING[target_lang]
|
351 |
translated_text = "Translation not available"
|
|
|
352 |
if model_status["mt"] == "loaded" and mt_model is not None and mt_tokenizer is not None:
|
353 |
try:
|
354 |
source_nllb_code = NLLB_LANGUAGE_CODES[source_code]
|
@@ -369,20 +422,26 @@ async def translate_text(text: str = Form(...), source_lang: str = Form(...), ta
|
|
369 |
translated_text = f"Translation failed: {str(e)}"
|
370 |
else:
|
371 |
logger.warning("MT model not loaded, skipping translation")
|
|
|
|
|
372 |
is_inappropriate = check_inappropriate_content(text) or check_inappropriate_content(translated_text)
|
373 |
if is_inappropriate:
|
374 |
logger.warning("Inappropriate content detected in translation request")
|
|
|
|
|
375 |
output_audio_url = None
|
376 |
if model_status["tts"].startswith("loaded"):
|
|
|
377 |
if load_tts_model_for_language(target_code):
|
378 |
try:
|
379 |
output_path, error = synthesize_speech(translated_text, target_code)
|
380 |
if output_path:
|
381 |
output_filename = os.path.basename(output_path)
|
382 |
-
output_audio_url = f"https://jerich-
|
383 |
logger.info("TTS conversion completed")
|
384 |
except Exception as e:
|
385 |
logger.error(f"Error during TTS conversion: {str(e)}")
|
|
|
386 |
return {
|
387 |
"request_id": request_id,
|
388 |
"status": "completed",
|
@@ -395,7 +454,8 @@ async def translate_text(text: str = Form(...), source_lang: str = Form(...), ta
|
|
395 |
|
396 |
@app.post("/translate-audio")
|
397 |
async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form(...), target_lang: str = Form(...)):
|
398 |
-
|
|
|
399 |
global mt_model, mt_tokenizer, tts_model, tts_tokenizer, current_tts_language
|
400 |
|
401 |
if not audio:
|
@@ -403,15 +463,19 @@ async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form
|
|
403 |
if source_lang not in LANGUAGE_MAPPING or target_lang not in LANGUAGE_MAPPING:
|
404 |
raise HTTPException(status_code=400, detail="Invalid language selected")
|
405 |
|
406 |
-
|
|
|
|
|
|
|
407 |
request_id = str(uuid.uuid4())
|
408 |
|
409 |
-
|
410 |
-
use_whisper = source_code in
|
411 |
-
|
412 |
-
|
413 |
-
|
414 |
-
|
|
|
415 |
return {
|
416 |
"request_id": request_id,
|
417 |
"status": "processing",
|
@@ -421,8 +485,9 @@ async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form
|
|
421 |
"output_audio": None,
|
422 |
"is_inappropriate": False
|
423 |
}
|
424 |
-
|
425 |
-
|
|
|
426 |
return {
|
427 |
"request_id": request_id,
|
428 |
"status": "processing",
|
@@ -433,6 +498,7 @@ async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form
|
|
433 |
"is_inappropriate": False
|
434 |
}
|
435 |
|
|
|
436 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_file:
|
437 |
temp_file.write(await audio.read())
|
438 |
temp_path = temp_file.name
|
@@ -443,16 +509,19 @@ async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form
|
|
443 |
is_inappropriate = False
|
444 |
|
445 |
try:
|
|
|
446 |
logger.info(f"Reading audio file: {temp_path}")
|
447 |
waveform, sample_rate = torchaudio.load(temp_path)
|
448 |
logger.info(f"Audio loaded: sample_rate={sample_rate}, waveform_shape={waveform.shape}")
|
449 |
|
|
|
450 |
if sample_rate != 16000:
|
451 |
logger.info(f"Resampling audio from {sample_rate} Hz to 16000 Hz")
|
452 |
resampler = torchaudio.transforms.Resample(sample_rate, 16000)
|
453 |
waveform = resampler(waveform)
|
454 |
sample_rate = 16000
|
455 |
|
|
|
456 |
if not detect_speech(waveform, sample_rate):
|
457 |
return {
|
458 |
"request_id": request_id,
|
@@ -464,26 +533,45 @@ async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form
|
|
464 |
"is_inappropriate": False
|
465 |
}
|
466 |
|
|
|
467 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
468 |
-
logger.info(f"Using device: {device}")
|
469 |
|
470 |
if use_whisper:
|
471 |
-
|
472 |
-
|
473 |
-
|
|
|
|
|
|
|
|
|
474 |
with torch.no_grad():
|
475 |
-
|
476 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
477 |
else:
|
478 |
-
|
479 |
-
|
|
|
|
|
|
|
|
|
|
|
480 |
with torch.no_grad():
|
481 |
-
|
|
|
482 |
predicted_ids = torch.argmax(logits, dim=-1)
|
483 |
-
transcription =
|
|
|
484 |
logger.info(f"Transcription completed: {transcription}")
|
485 |
|
486 |
-
|
487 |
if model_status["mt"] == "loaded" and mt_model is not None and mt_tokenizer is not None:
|
488 |
try:
|
489 |
source_nllb_code = NLLB_LANGUAGE_CODES[source_code]
|
@@ -504,16 +592,18 @@ async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form
|
|
504 |
else:
|
505 |
logger.warning("MT model not loaded, skipping translation")
|
506 |
|
|
|
507 |
is_inappropriate = check_inappropriate_content(transcription) or check_inappropriate_content(translated_text)
|
508 |
if is_inappropriate:
|
509 |
logger.warning("Inappropriate content detected in audio transcription or translation")
|
510 |
|
|
|
511 |
if load_tts_model_for_language(target_code):
|
512 |
try:
|
513 |
output_path, error = synthesize_speech(translated_text, target_code)
|
514 |
if output_path:
|
515 |
output_filename = os.path.basename(output_path)
|
516 |
-
output_audio_url = f"https://jerich-
|
517 |
logger.info("TTS conversion completed")
|
518 |
except Exception as e:
|
519 |
logger.error(f"Error during TTS conversion: {str(e)}")
|
@@ -544,6 +634,7 @@ async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form
|
|
544 |
|
545 |
@app.post("/text-to-speech")
|
546 |
async def text_to_speech(text: str = Form(...), target_lang: str = Form(...)):
|
|
|
547 |
if not text:
|
548 |
raise HTTPException(status_code=400, detail="No text provided")
|
549 |
if target_lang not in LANGUAGE_MAPPING:
|
@@ -553,17 +644,20 @@ async def text_to_speech(text: str = Form(...), target_lang: str = Form(...)):
|
|
553 |
request_id = str(uuid.uuid4())
|
554 |
|
555 |
target_code = LANGUAGE_MAPPING[target_lang]
|
|
|
|
|
556 |
is_inappropriate = check_inappropriate_content(text)
|
557 |
if is_inappropriate:
|
558 |
logger.warning("Inappropriate content detected in text-to-speech request")
|
559 |
|
|
|
560 |
output_audio_url = None
|
561 |
if model_status["tts"].startswith("loaded") or load_tts_model_for_language(target_code):
|
562 |
try:
|
563 |
output_path, error = synthesize_speech(text, target_code)
|
564 |
if output_path:
|
565 |
output_filename = os.path.basename(output_path)
|
566 |
-
output_audio_url = f"https://jerich-
|
567 |
logger.info("TTS conversion completed")
|
568 |
else:
|
569 |
logger.error(f"TTS conversion failed: {error}")
|
|
|
17 |
from fastapi.responses import JSONResponse
|
18 |
from fastapi.staticfiles import StaticFiles
|
19 |
from typing import Dict, Any, Optional, Tuple, List
|
20 |
+
from datetime import datetime, timedelta
|
21 |
|
22 |
# Configure logging
|
23 |
logging.basicConfig(level=logging.INFO)
|
|
|
35 |
loading_in_progress = False
|
36 |
loading_thread = None
|
37 |
model_status = {
|
38 |
+
"stt_whisper": "not_loaded",
|
39 |
+
"stt_mms": "not_loaded",
|
40 |
"mt": "not_loaded",
|
41 |
"tts": "not_loaded"
|
42 |
}
|
|
|
44 |
current_tts_language = "tgl" # Track the current TTS language
|
45 |
|
46 |
# Model instances
|
47 |
+
whisper_processor = None
|
48 |
+
whisper_model = None
|
49 |
+
mms_processor = None
|
50 |
+
mms_model = None
|
51 |
mt_model = None
|
52 |
mt_tokenizer = None
|
53 |
tts_model = None
|
|
|
63 |
"Pangasinan": "pag"
|
64 |
}
|
65 |
|
66 |
+
# Define which languages use Whisper vs MMS for STT
|
67 |
+
WHISPER_LANGUAGES = {"eng", "tgl"} # English and Tagalog use Whisper
|
68 |
+
MMS_LANGUAGES = {"ceb", "ilo", "war", "pag"} # Other Philippine languages use MMS
|
|
|
|
|
69 |
|
70 |
NLLB_LANGUAGE_CODES = {
|
71 |
"eng": "eng_Latn",
|
|
|
92 |
Check if the text contains inappropriate content.
|
93 |
Returns True if inappropriate content is detected, False otherwise.
|
94 |
"""
|
95 |
+
# Convert to lowercase for case-insensitive matching
|
96 |
text_lower = text.lower()
|
97 |
+
|
98 |
+
# Check for inappropriate words
|
99 |
for word in INAPPROPRIATE_WORDS:
|
100 |
+
# Use word boundary matching to avoid false positives
|
101 |
pattern = r'\b' + re.escape(word) + r'\b'
|
102 |
if re.search(pattern, text_lower):
|
103 |
logger.warning(f"Inappropriate content detected: {word}")
|
104 |
return True
|
105 |
+
|
106 |
return False
|
107 |
|
108 |
# Function to save PCM data as a WAV file
|
109 |
def save_pcm_to_wav(pcm_data: list, sample_rate: int, output_path: str):
|
110 |
+
# Convert pcm_data to a NumPy array of 16-bit integers
|
111 |
pcm_array = np.array(pcm_data, dtype=np.int16)
|
112 |
+
|
113 |
with wave.open(output_path, 'wb') as wav_file:
|
114 |
+
# Set WAV parameters: 1 channel (mono), 2 bytes per sample (16-bit), sample rate
|
115 |
wav_file.setnchannels(1)
|
116 |
+
wav_file.setsampwidth(2) # 16-bit audio
|
117 |
wav_file.setframerate(sample_rate)
|
118 |
+
# Write the 16-bit PCM data as bytes (little-endian)
|
119 |
wav_file.writeframes(pcm_array.tobytes())
|
120 |
|
121 |
# Function to detect speech using an energy-based approach
|
122 |
def detect_speech(waveform: torch.Tensor, sample_rate: int, threshold: float = 0.01, min_speech_duration: float = 0.5) -> bool:
|
123 |
+
"""
|
124 |
+
Detects if the audio contains speech using an energy-based approach.
|
125 |
+
Returns True if speech is detected, False otherwise.
|
126 |
+
"""
|
127 |
+
# Convert waveform to numpy array
|
128 |
waveform_np = waveform.numpy()
|
129 |
if waveform_np.ndim > 1:
|
130 |
+
waveform_np = waveform_np.mean(axis=0) # Convert stereo to mono
|
131 |
+
|
132 |
+
# Compute RMS energy
|
133 |
rms = np.sqrt(np.mean(waveform_np**2))
|
134 |
logger.info(f"RMS energy: {rms}")
|
135 |
+
|
136 |
+
# Check if RMS energy exceeds the threshold
|
137 |
if rms < threshold:
|
138 |
logger.info("No speech detected: RMS energy below threshold")
|
139 |
return False
|
140 |
+
|
141 |
+
# Optionally, check for minimum speech duration (requires more sophisticated VAD)
|
142 |
+
# For now, we assume if RMS is above threshold, there is speech
|
143 |
return True
|
144 |
|
145 |
# Function to clean up old audio files
|
146 |
def cleanup_old_audio_files():
|
147 |
logger.info("Starting cleanup of old audio files...")
|
148 |
+
expiration_time = datetime.now() - timedelta(minutes=10) # Files older than 10 minutes
|
149 |
for filename in os.listdir(AUDIO_DIR):
|
150 |
file_path = os.path.join(AUDIO_DIR, filename)
|
151 |
if os.path.isfile(file_path):
|
|
|
161 |
def schedule_cleanup():
|
162 |
while True:
|
163 |
cleanup_old_audio_files()
|
164 |
+
time.sleep(300) # Run every 5 minutes (300 seconds)
|
165 |
|
166 |
# Function to load models in background
|
167 |
def load_models_task():
|
168 |
global models_loaded, loading_in_progress, model_status, error_message
|
169 |
+
global whisper_processor, whisper_model, mms_processor, mms_model
|
170 |
global mt_model, mt_tokenizer, tts_model, tts_tokenizer
|
171 |
|
172 |
try:
|
173 |
loading_in_progress = True
|
174 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
175 |
|
176 |
+
# Load Whisper STT model for English and Tagalog
|
177 |
+
logger.info("Starting to load Whisper STT model...")
|
178 |
+
from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
179 |
|
180 |
try:
|
181 |
logger.info("Loading Whisper STT model...")
|
182 |
+
model_status["stt_whisper"] = "loading"
|
183 |
+
whisper_processor = WhisperProcessor.from_pretrained("openai/whisper-small")
|
184 |
+
whisper_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
|
185 |
+
whisper_model.to(device)
|
|
|
186 |
logger.info("Whisper STT model loaded successfully")
|
187 |
+
model_status["stt_whisper"] = "loaded"
|
188 |
+
except Exception as whisper_error:
|
189 |
+
logger.error(f"Failed to load Whisper STT model: {str(whisper_error)}")
|
190 |
+
model_status["stt_whisper"] = "failed"
|
191 |
+
error_message = f"Whisper STT model loading failed: {str(whisper_error)}"
|
192 |
return
|
193 |
+
|
194 |
+
# Load MMS STT model for other Philippine languages
|
195 |
+
logger.info("Starting to load MMS STT model...")
|
196 |
+
from transformers import AutoProcessor, AutoModelForCTC
|
197 |
+
|
198 |
try:
|
199 |
logger.info("Loading MMS STT model...")
|
200 |
+
model_status["stt_mms"] = "loading"
|
201 |
+
mms_processor = AutoProcessor.from_pretrained("facebook/mms-1b-all")
|
202 |
+
mms_model = AutoModelForCTC.from_pretrained("facebook/mms-1b-all")
|
203 |
+
mms_model.to(device)
|
204 |
logger.info("MMS STT model loaded successfully")
|
205 |
+
model_status["stt_mms"] = "loaded"
|
206 |
+
except Exception as mms_error:
|
207 |
+
logger.error(f"Failed to load MMS STT model: {str(mms_error)}")
|
208 |
+
model_status["stt_mms"] = "failed"
|
209 |
+
error_message = f"MMS STT model loading failed: {str(mms_error)}"
|
210 |
+
return
|
|
|
211 |
|
212 |
# Load MT model
|
213 |
logger.info("Starting to load MT model...")
|
|
|
227 |
error_message = f"MT model loading failed: {str(e)}"
|
228 |
return
|
229 |
|
230 |
+
# Load TTS model (default to Tagalog, will be updated dynamically)
|
231 |
logger.info("Starting to load TTS model...")
|
232 |
from transformers import VitsModel, AutoTokenizer
|
233 |
|
|
|
241 |
model_status["tts"] = "loaded"
|
242 |
except Exception as e:
|
243 |
logger.error(f"Failed to load TTS model for Tagalog: {str(e)}")
|
244 |
+
# Fallback to English TTS if the target language fails
|
245 |
try:
|
246 |
logger.info("Falling back to MMS-TTS English model...")
|
247 |
tts_model = VitsModel.from_pretrained("facebook/mms-tts-eng")
|
|
|
282 |
|
283 |
# Function to load or update TTS model for a specific language
|
284 |
def load_tts_model_for_language(target_code: str) -> bool:
|
285 |
+
"""
|
286 |
+
Load or update the TTS model for the specified language.
|
287 |
+
Returns True if successful, False otherwise.
|
288 |
+
"""
|
289 |
global tts_model, tts_tokenizer, current_tts_language, model_status
|
290 |
+
|
291 |
if target_code not in LANGUAGE_MAPPING.values():
|
292 |
logger.error(f"Invalid language code: {target_code}")
|
293 |
return False
|
294 |
+
|
295 |
+
# Skip if the model is already loaded for the target language
|
296 |
if current_tts_language == target_code and model_status["tts"].startswith("loaded"):
|
297 |
logger.info(f"TTS model for {target_code} is already loaded.")
|
298 |
return True
|
299 |
+
|
300 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
301 |
try:
|
302 |
logger.info(f"Loading MMS-TTS model for {target_code}...")
|
|
|
326 |
|
327 |
# Function to synthesize speech from text
|
328 |
def synthesize_speech(text: str, target_code: str) -> Tuple[Optional[str], Optional[str]]:
|
329 |
+
"""
|
330 |
+
Convert text to speech for the specified language.
|
331 |
+
Returns a tuple of (output_path, error_message).
|
332 |
+
"""
|
333 |
global tts_model, tts_tokenizer
|
334 |
+
|
335 |
request_id = str(uuid.uuid4())
|
336 |
output_path = os.path.join(AUDIO_DIR, f"{request_id}.wav")
|
337 |
+
|
338 |
+
# Make sure the TTS model is loaded for the target language
|
339 |
if not load_tts_model_for_language(target_code):
|
340 |
return None, "Failed to load TTS model for the target language"
|
341 |
+
|
342 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
343 |
try:
|
344 |
+
inputs = tts_tokenizer(text, return_tensors="pt").to(device)
|
345 |
with torch.no_grad():
|
346 |
output = tts_model(**inputs)
|
347 |
speech = output.waveform.cpu().numpy().squeeze()
|
348 |
speech = (speech * 32767).astype(np.int16)
|
349 |
sample_rate = tts_model.config.sampling_rate
|
350 |
+
|
351 |
+
# Save the audio as a WAV file
|
352 |
save_pcm_to_wav(speech.tolist(), sample_rate, output_path)
|
353 |
logger.info(f"Saved synthesized audio to {output_path}")
|
354 |
+
|
355 |
return output_path, None
|
356 |
except Exception as e:
|
357 |
error_msg = f"Error during TTS conversion: {str(e)}"
|
|
|
367 |
|
368 |
@app.get("/")
|
369 |
async def root():
|
370 |
+
"""Root endpoint for default health check"""
|
371 |
logger.info("Root endpoint requested")
|
372 |
return {"status": "healthy"}
|
373 |
|
374 |
@app.get("/health")
|
375 |
async def health_check():
|
376 |
+
"""Health check endpoint that always returns successfully"""
|
377 |
+
global models_loaded, loading_in_progress, model_status, error_message
|
378 |
logger.info("Health check requested")
|
379 |
return {
|
380 |
"status": "healthy",
|
|
|
386 |
|
387 |
@app.post("/translate-text")
|
388 |
async def translate_text(text: str = Form(...), source_lang: str = Form(...), target_lang: str = Form(...)):
|
389 |
+
"""Endpoint to translate text and convert to speech"""
|
390 |
global mt_model, mt_tokenizer, tts_model, tts_tokenizer, current_tts_language
|
391 |
+
|
392 |
if not text:
|
393 |
raise HTTPException(status_code=400, detail="No text provided")
|
394 |
if source_lang not in LANGUAGE_MAPPING or target_lang not in LANGUAGE_MAPPING:
|
395 |
raise HTTPException(status_code=400, detail="Invalid language selected")
|
396 |
+
|
397 |
logger.info(f"Translate-text requested: {text} from {source_lang} to {target_lang}")
|
398 |
request_id = str(uuid.uuid4())
|
399 |
+
|
400 |
+
# Translate the text
|
401 |
source_code = LANGUAGE_MAPPING[source_lang]
|
402 |
target_code = LANGUAGE_MAPPING[target_lang]
|
403 |
translated_text = "Translation not available"
|
404 |
+
|
405 |
if model_status["mt"] == "loaded" and mt_model is not None and mt_tokenizer is not None:
|
406 |
try:
|
407 |
source_nllb_code = NLLB_LANGUAGE_CODES[source_code]
|
|
|
422 |
translated_text = f"Translation failed: {str(e)}"
|
423 |
else:
|
424 |
logger.warning("MT model not loaded, skipping translation")
|
425 |
+
|
426 |
+
# Check for inappropriate content in the source text and translated text
|
427 |
is_inappropriate = check_inappropriate_content(text) or check_inappropriate_content(translated_text)
|
428 |
if is_inappropriate:
|
429 |
logger.warning("Inappropriate content detected in translation request")
|
430 |
+
|
431 |
+
# Convert translated text to speech
|
432 |
output_audio_url = None
|
433 |
if model_status["tts"].startswith("loaded"):
|
434 |
+
# Load or update TTS model for the target language
|
435 |
if load_tts_model_for_language(target_code):
|
436 |
try:
|
437 |
output_path, error = synthesize_speech(translated_text, target_code)
|
438 |
if output_path:
|
439 |
output_filename = os.path.basename(output_path)
|
440 |
+
output_audio_url = f"https://jerich-talklasapp.hf.space/audio_output/{output_filename}"
|
441 |
logger.info("TTS conversion completed")
|
442 |
except Exception as e:
|
443 |
logger.error(f"Error during TTS conversion: {str(e)}")
|
444 |
+
|
445 |
return {
|
446 |
"request_id": request_id,
|
447 |
"status": "completed",
|
|
|
454 |
|
455 |
@app.post("/translate-audio")
|
456 |
async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form(...), target_lang: str = Form(...)):
|
457 |
+
"""Endpoint to transcribe, translate, and convert audio to speech"""
|
458 |
+
global whisper_processor, whisper_model, mms_processor, mms_model
|
459 |
global mt_model, mt_tokenizer, tts_model, tts_tokenizer, current_tts_language
|
460 |
|
461 |
if not audio:
|
|
|
463 |
if source_lang not in LANGUAGE_MAPPING or target_lang not in LANGUAGE_MAPPING:
|
464 |
raise HTTPException(status_code=400, detail="Invalid language selected")
|
465 |
|
466 |
+
source_code = LANGUAGE_MAPPING[source_lang]
|
467 |
+
target_code = LANGUAGE_MAPPING[target_lang]
|
468 |
+
|
469 |
+
logger.info(f"Translate-audio requested: {audio.filename} from {source_lang} ({source_code}) to {target_lang} ({target_code})")
|
470 |
request_id = str(uuid.uuid4())
|
471 |
|
472 |
+
# Determine which STT model to use based on source language
|
473 |
+
use_whisper = source_code in WHISPER_LANGUAGES
|
474 |
+
use_mms = source_code in MMS_LANGUAGES
|
475 |
+
|
476 |
+
# Check if the appropriate STT model is loaded
|
477 |
+
if use_whisper and (model_status["stt_whisper"] != "loaded" or whisper_processor is None or whisper_model is None):
|
478 |
+
logger.warning("Whisper STT model not loaded for English/Tagalog, returning placeholder response")
|
479 |
return {
|
480 |
"request_id": request_id,
|
481 |
"status": "processing",
|
|
|
485 |
"output_audio": None,
|
486 |
"is_inappropriate": False
|
487 |
}
|
488 |
+
|
489 |
+
if use_mms and (model_status["stt_mms"] != "loaded" or mms_processor is None or mms_model is None):
|
490 |
+
logger.warning("MMS STT model not loaded for Philippine languages, returning placeholder response")
|
491 |
return {
|
492 |
"request_id": request_id,
|
493 |
"status": "processing",
|
|
|
498 |
"is_inappropriate": False
|
499 |
}
|
500 |
|
501 |
+
# Save the uploaded audio to a temporary file
|
502 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_file:
|
503 |
temp_file.write(await audio.read())
|
504 |
temp_path = temp_file.name
|
|
|
509 |
is_inappropriate = False
|
510 |
|
511 |
try:
|
512 |
+
# Step 1: Load and resample the audio using torchaudio
|
513 |
logger.info(f"Reading audio file: {temp_path}")
|
514 |
waveform, sample_rate = torchaudio.load(temp_path)
|
515 |
logger.info(f"Audio loaded: sample_rate={sample_rate}, waveform_shape={waveform.shape}")
|
516 |
|
517 |
+
# Resample to 16 kHz if needed (required by Whisper and MMS models)
|
518 |
if sample_rate != 16000:
|
519 |
logger.info(f"Resampling audio from {sample_rate} Hz to 16000 Hz")
|
520 |
resampler = torchaudio.transforms.Resample(sample_rate, 16000)
|
521 |
waveform = resampler(waveform)
|
522 |
sample_rate = 16000
|
523 |
|
524 |
+
# Step 2: Detect speech
|
525 |
if not detect_speech(waveform, sample_rate):
|
526 |
return {
|
527 |
"request_id": request_id,
|
|
|
533 |
"is_inappropriate": False
|
534 |
}
|
535 |
|
536 |
+
# Step 3: Transcribe the audio (STT)
|
537 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
538 |
+
logger.info(f"Using device: {device} for STT")
|
539 |
|
540 |
if use_whisper:
|
541 |
+
# Use Whisper model for English and Tagalog
|
542 |
+
logger.info(f"Using Whisper model for language: {source_code}")
|
543 |
+
|
544 |
+
# Prepare audio for Whisper
|
545 |
+
inputs = whisper_processor(waveform.numpy(), sampling_rate=16000, return_tensors="pt").to(device)
|
546 |
+
logger.info("Audio processed for Whisper, generating transcription...")
|
547 |
+
|
548 |
with torch.no_grad():
|
549 |
+
# For English, we can specify the language; for Tagalog we use 'tl'
|
550 |
+
forced_language = "en" if source_code == "eng" else "tl"
|
551 |
+
generated_ids = whisper_model.generate(
|
552 |
+
**inputs,
|
553 |
+
language=forced_language,
|
554 |
+
task="transcribe"
|
555 |
+
)
|
556 |
+
transcription = whisper_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
557 |
+
|
558 |
else:
|
559 |
+
# Use MMS model for other Philippine languages
|
560 |
+
logger.info(f"Using MMS model for language: {source_code}")
|
561 |
+
|
562 |
+
# Prepare audio for MMS
|
563 |
+
inputs = mms_processor(waveform.numpy(), sampling_rate=16000, return_tensors="pt").to(device)
|
564 |
+
logger.info("Audio processed for MMS, generating transcription...")
|
565 |
+
|
566 |
with torch.no_grad():
|
567 |
+
# Process with MMS
|
568 |
+
logits = mms_model(**inputs).logits
|
569 |
predicted_ids = torch.argmax(logits, dim=-1)
|
570 |
+
transcription = mms_processor.batch_decode(predicted_ids)[0]
|
571 |
+
|
572 |
logger.info(f"Transcription completed: {transcription}")
|
573 |
|
574 |
+
# Step 4: Translate the transcribed text (MT)
|
575 |
if model_status["mt"] == "loaded" and mt_model is not None and mt_tokenizer is not None:
|
576 |
try:
|
577 |
source_nllb_code = NLLB_LANGUAGE_CODES[source_code]
|
|
|
592 |
else:
|
593 |
logger.warning("MT model not loaded, skipping translation")
|
594 |
|
595 |
+
# Step 5: Check for inappropriate content
|
596 |
is_inappropriate = check_inappropriate_content(transcription) or check_inappropriate_content(translated_text)
|
597 |
if is_inappropriate:
|
598 |
logger.warning("Inappropriate content detected in audio transcription or translation")
|
599 |
|
600 |
+
# Step 6: Convert translated text to speech (TTS)
|
601 |
if load_tts_model_for_language(target_code):
|
602 |
try:
|
603 |
output_path, error = synthesize_speech(translated_text, target_code)
|
604 |
if output_path:
|
605 |
output_filename = os.path.basename(output_path)
|
606 |
+
output_audio_url = f"https://jerich-talklasapp.hf.space/audio_output/{output_filename}"
|
607 |
logger.info("TTS conversion completed")
|
608 |
except Exception as e:
|
609 |
logger.error(f"Error during TTS conversion: {str(e)}")
|
|
|
634 |
|
635 |
@app.post("/text-to-speech")
|
636 |
async def text_to_speech(text: str = Form(...), target_lang: str = Form(...)):
|
637 |
+
"""Endpoint to convert text to speech in the specified language"""
|
638 |
if not text:
|
639 |
raise HTTPException(status_code=400, detail="No text provided")
|
640 |
if target_lang not in LANGUAGE_MAPPING:
|
|
|
644 |
request_id = str(uuid.uuid4())
|
645 |
|
646 |
target_code = LANGUAGE_MAPPING[target_lang]
|
647 |
+
|
648 |
+
# Check for inappropriate content
|
649 |
is_inappropriate = check_inappropriate_content(text)
|
650 |
if is_inappropriate:
|
651 |
logger.warning("Inappropriate content detected in text-to-speech request")
|
652 |
|
653 |
+
# Synthesize speech
|
654 |
output_audio_url = None
|
655 |
if model_status["tts"].startswith("loaded") or load_tts_model_for_language(target_code):
|
656 |
try:
|
657 |
output_path, error = synthesize_speech(text, target_code)
|
658 |
if output_path:
|
659 |
output_filename = os.path.basename(output_path)
|
660 |
+
output_audio_url = f"https://jerich-talklasapp.hf.space/audio_output/{output_filename}"
|
661 |
logger.info("TTS conversion completed")
|
662 |
else:
|
663 |
logger.error(f"TTS conversion failed: {error}")
|