Update app.py

app.py

@@ -1,13 +1,3 @@
-# app.py — CareCompanion: Alzheimer's Caregiver Voice Assistant
-#
-# Stack:
-#   STT:  distil-whisper/distil-large-v3   (local, fast)
-#   LLM:  openai/gpt-oss-20b + FAISS RAG   (HF Inference API)
-#   TTS:  parler-tts/parler-tts-mini-v1    (local, neutral American voice)
-#
-# Secrets needed in HF Space Settings:
-#   HF_TOKEN — your Hugging Face access token
-
 import os
 import faiss
 import pickle
@@ -20,7 +10,7 @@ import tempfile
 from huggingface_hub import hf_hub_download
 from sentence_transformers import SentenceTransformer
 from huggingface_hub import InferenceClient
-from transformers import VitsModel, AutoTokenizer, pipeline, SpeechT5HifiGan
+from transformers import VitsModel, AutoTokenizer, pipeline
 
 
 # ── Auth ───────────────────────────────────────────────────────
@@ -95,59 +85,58 @@ def detect_language(text):
     try:
         lang = detect(text)
         if lang == "ca":
-            return "Català"
+            return "ca"
         elif lang == "es":
-            return "Español"
+            return "es"
+        elif lang == "en":
+            return "en"
         else:
-            return "Español"
+            return "es"
     except:
-        return "Español"
+        return "es"
         
 # ── TTS: Parler TTS mini v1 (neutral català/spanish voice) ─────────
-#print("Loading MMS TTS models...")
-#tts_models, tts_tokenizers = {}, {}
-#for lang_code, repo in {"en": "facebook/mms-tts-eng", "es": "facebook/mms-tts-spa", "ca": "facebook/mms-tts-cat"}.items():
-#    tts_tokenizers[lang_code] = AutoTokenizer.from_pretrained(repo)
-#    tts_models[lang_code] = VitsModel.from_pretrained(repo).to(device)
-#    tts_models[lang_code].eval()
+print("Loading MMS TTS models...")
+tts_models, tts_tokenizers = {}, {}
+for lang_code, repo in {"en": "facebook/mms-tts-eng", "es": "facebook/mms-tts-spa", "ca": "facebook/mms-tts-cat"}.items():
+    tts_tokenizers[lang_code] = AutoTokenizer.from_pretrained(repo)
+    tts_models[lang_code] = VitsModel.from_pretrained(repo).to(device)
+    tts_models[lang_code].eval()
 print("Loading TTS models...")
 
-# Kokoro for English and Spanish
-from kokoro import KPipeline
-kokoro_en = KPipeline(lang_code='en')
-kokoro_es = KPipeline(lang_code='es')
-
-# Matxa (BSC) for Catalan
-tts_tokenizers, tts_models = {}, {}
-tts_tokenizers["ca"] = AutoTokenizer.from_pretrained("projecte-aina/matxa-tts-cat-multiaccent")
-tts_models["ca"] = VitsModel.from_pretrained("projecte-aina/matxa-tts-cat-multiaccent").to(device)
-tts_models["ca"].eval()
-
 def text_to_speech(text, lang="es"):
-    if not text:
+    if not text or lang not in tts_models:
         return None
     try:
-        if lang == "ca":
-            inputs = tts_tokenizers["ca"](text, return_tensors="pt").to(device)
-            with torch.no_grad():
-                audio = tts_models["ca"](**inputs).waveform
-            audio_int16 = (audio.squeeze().cpu().float().numpy() * 32767).clip(-32768, 32767).astype("int16")
-            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
-                scipy.io.wavfile.write(f.name, rate=tts_models["ca"].config.sampling_rate, data=audio_int16)
-                return f.name
-        else:
-            pipeline = kokoro_en if lang == "en" else kokoro_es
-            voice = "af_heart" if lang == "en" else "ef_dora"
-            audio_chunks = []
-            for _, _, audio in pipeline(text, voice=voice):
-                audio_chunks.append(audio)
-            audio_np = np.concatenate(audio_chunks)
-            audio_int16 = (audio_np * 32767).clip(-32768, 32767).astype("int16")
-            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
-                scipy.io.wavfile.write(f.name, rate=24000, data=audio_int16)
-                return f.name
+        inputs = tts_tokenizers[lang](text, return_tensors="pt").to(device)
+        with torch.no_grad():
+            audio = tts_models[lang](**inputs).waveform
+        audio_int16 = (audio.squeeze().cpu().float().numpy() * 32767).clip(-32768, 32767).astype("int16")
+        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
+            scipy.io.wavfile.write(f.name, rate=tts_models[lang].config.sampling_rate, data=audio_int16)
+            return f.name
+    except Exception as e:
+        print(f"TTS error: {e}")
+        return None
+
+    try:
+        input_ids = tts_tokenizer(voice_desc, return_tensors="pt").input_ids.to(device)
+        prompt_ids = tts_tokenizer(text, return_tensors="pt").input_ids.to(device)
+
+        with torch.no_grad():
+            generation = tts_model.generate(
+                input_ids=input_ids,
+                prompt_input_ids=prompt_ids,
+            )
+
+        audio_array = generation.cpu().to(torch.float32).numpy().squeeze()
+
+        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
+            scipy.io.wavfile.write(f.name, rate=sampling_rate, data=audio_array)
+            return f.name
+
     except Exception as e:
-        print(f"TTS error ({lang}): {e}")
+        print(f"TTS error: {e}")
         return None
 
 
@@ -165,7 +154,7 @@ If asked about local resources, ONLY reference services mentioned in the retriev
 If no relevant local services are in the context, say so honestly.
 Always remind caregivers that asking for help is a sign of strength, not weakness."""
 
-def respond_to_message(message, history, lang="Español"):
+def respond_to_message(message, history, lang="es"):
     if not message.strip():
         return ""