Update utils_hf.py

utils_hf.py

@@ -14,7 +14,6 @@ from transformers import pipeline
 from openai import Client
 from dotenv import load_dotenv
 
-
 load_dotenv()
 
 hf_token = os.environ.get("HF_DIARIZATION_TOKEN")
@@ -33,12 +32,12 @@ if torch.cuda.is_available():
     pyannote_pipeline.to(torch.device("cuda"))
 
 sumamry_pipeline = pipeline(
-    "summarization",
-    model="t5-base",
+    "text-generation",
+    model="gpt-3.5-turbo",
     device=device,
 )
 
-whisper_model = whisper.load_model("base")
+whisper_model = whisper.load_model("turbo")
 
 
 def transcribir_segmentos(diarization_segments, audio_path):
@@ -63,7 +62,7 @@ def transcribir_segmentos(diarization_segments, audio_path):
             segment_audio_path = f"segment_{start_time}_{end_time}.wav"
             extract_audio_segment(audio_path, segment_audio_path, start_time, end_time)
 
-            transcript = transcribe_audio_whisper_lib(segment_audio_path) # Cambiar aca el trancript
+            transcript = transcribe_audio_whisper_lib(segment_audio_path)  # Cambiar aca el trancript
             if transcript:
                 diarization_text += f"{speaker}: {transcript}\n"
                 file.write(f"{speaker}: {transcript}\n")
@@ -93,7 +92,9 @@ def summarize_speaker_transcripts(speaker_transcripts):
         if len(full_text) > max_tokens:
             full_text = full_text[:max_tokens]
 
-        summary = sumamry_pipeline(full_text, max_length=150, min_length=30, do_sample=False)[0]['summary_text']
+        prompt = f"Por favor, resume el siguiente texto:\n\n{full_text}"
+
+        summary = sumamry_pipeline(prompt, max_length=150, min_length=30, do_sample=False)[0]['text']
         summaries[speaker] = summary
     return summaries
 
@@ -124,8 +125,6 @@ def diarize_full_audio(audio_path):
         return None, None, 0
 
 
-
-
 def extract_audio_segment(input_audio_path, output_audio_path, start_time, end_time):
     try:
         audio = AudioFileClip(input_audio_path)
@@ -150,7 +149,6 @@ def transcribe_audio_whisper_lib(audio_path):
         st.stop()
 
 
-
 def download_youtube_audio(url, progress_bar):
     # No incluyas la extensión '.wav' en el 'outtmpl'
     output_path = os.path.join('temp_audio')
@@ -198,6 +196,7 @@ def get_audio_duration(audio_path):
         st.error(f"Error al obtener la duración del audio: {e}")
         return 0
 
+
 def update_progress(d, progress_bar):
     if d['status'] == 'downloading':
         total_bytes = d.get('total_bytes', None)
@@ -214,6 +213,7 @@ def update_progress(d, progress_bar):
     elif d['status'] == 'postprocessing':
         progress_bar.progress(0.95)  # 95% durante el procesamiento de audio
 
+
 class StreamlitProgressHook(ProgressHook):
     def __init__(self, transient: bool = False):
         super().__init__(transient)
@@ -227,48 +227,3 @@ class StreamlitProgressHook(ProgressHook):
             progress_message = f"{step_name:<20} ━ Progress data unavailable"
         self.progress_text.text(progress_message)
 
-
-# def diarize_in_segments(audio_path, segment_duration=300):
-#     try:
-#         audio_duration = get_audio_duration(audio_path)
-#         diarization_segments = []
-#         diarization_text = ""
-#         speakers = set()
-#
-#         # Procesar cada segmento por separado
-#         for start_time in range(0, int(audio_duration), segment_duration):
-#             end_time = min(start_time + segment_duration, audio_duration)
-#             st.info(f"Procesando segmento desde {start_time} hasta {end_time} segundos...")
-#
-#             with StreamlitProgressHook() as hook:
-#                 diarization = pyannote_pipeline({
-#                     'audio': audio_path,
-#                     'start': start_time,
-#                     'end': end_time
-#                 }, hook=hook)
-#
-#             for segment, _, speaker in diarization.itertracks(yield_label=True):
-#                 diarization_segments.append((segment.start, segment.end, speaker))
-#                 diarization_text += f"{segment.start:.2f} - {segment.end:.2f}: {speaker}\n"
-#                 speakers.add(speaker)
-#
-#         return diarization_segments, diarization_text, speakers
-#     except Exception as e:
-#         st.error(f"Error al realizar la diarización de un segmento: {e}")
-#         return None, None, 0
-
-# def transcribe_audio_whisper_transformers(audio_path):
-#     try:
-#         # Leer el archivo de audio y convertirlo en un numpy ndarray
-#         audio_data, _ = sf.read(audio_path)
-#
-#         # Convertir a un solo canal si es necesario
-#         if len(audio_data.shape) > 1:
-#             audio_data = np.mean(audio_data, axis=1)
-#
-#         # Transcribir el audio usando el modelo Whisper
-#         transcript = whisper_pipeline(audio_data, return_timestamps=True)
-#         return transcript['text']
-#     except Exception as e:
-#         st.error(f"Error al transcribir el audio: {e}")
-#         st.stop()