Upload app.py with huggingface_hub

app.py

@@ -39,9 +39,10 @@ os.environ["TRANSFORMERS_CACHE"] = str(HF_CACHE_DIR)
 
 F5_SPANISH_MODEL_ID = "jpgallegoar/F5-Spanish"
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-APP_VERSION = "1.0.1"
+APP_VERSION = "1.1.0"
 
 _f5_model = None
+_ref_text_cache = {}  # {audio_path: transcribed_text}
 
 
 def _clear_cache():
@@ -89,6 +90,18 @@ def _get_reference_audio():
     raise FileNotFoundError("No hay reference.wav. Descarga el modelo primero.")
 
 
+def _get_ref_text(audio_path):
+    """Pre-transcribe reference audio in Spanish to avoid Whisper auto-detecting wrong language."""
+    if audio_path in _ref_text_cache:
+        return _ref_text_cache[audio_path]
+    _load_tts()
+    logger.info(f"Transcribing reference audio as Spanish: {audio_path}")
+    ref_text = _f5_model.transcribe(audio_path, language="spanish")
+    logger.info(f"Reference transcription: {ref_text}")
+    _ref_text_cache[audio_path] = ref_text
+    return ref_text
+
+
 # ── Gradio handlers ──
 
 def download_model(project_name, progress=gr.Progress()):
@@ -135,6 +148,7 @@ def generate_speech(project_name, text, speed, progress=gr.Progress()):
         _load_tts()
 
         ref_audio = _get_reference_audio()
+        ref_text = _get_ref_text(ref_audio)
         output_path = str(TEMP_DIR / "tts_output.wav")
 
         progress(0.3, desc="Generando voz...")
@@ -142,7 +156,7 @@ def generate_speech(project_name, text, speed, progress=gr.Progress()):
 
         audio, sr, _spec = _f5_model.infer(
             ref_file=ref_audio,
-            ref_text="",
+            ref_text=ref_text,
             gen_text=text,
             speed=speed,
         )
@@ -174,9 +188,10 @@ def generate_with_custom_ref(project_name, text, ref_audio_path, speed, progress
         progress(0.3, desc="Generando voz...")
         logger.info(f"Generating with custom ref: '{text[:80]}...'")
 
+        ref_text = _get_ref_text(ref_audio_path)
         audio, sr, _spec = _f5_model.infer(
             ref_file=ref_audio_path,
-            ref_text="",
+            ref_text=ref_text,
             gen_text=text,
             speed=speed,
         )