Update app.py

app.py

@@ -68,15 +68,11 @@ def number_to_words(number):
         return str(number)
 
 def normalize_text(text):
-    # Remove commas from numbers like 1,000,000
     text = re.sub(r'(\d{1,3})(,\d{3})+', lambda m: m.group(0).replace(",", ""), text)
-    # Remove decimals (e.g., .00)
     text = re.sub(r'\.\d+', '', text)
-    # Replace numbers with Somali words
     def replace_num(match):
         return number_to_words(match.group())
     text = re.sub(r'\d+', replace_num, text)
-    # Replace special symbols
     symbol_map = {
         '$': 'doolar',
         '=': 'egwal',
@@ -85,7 +81,6 @@ def normalize_text(text):
     }
     for sym, word in symbol_map.items():
         text = text.replace(sym, ' ' + word + ' ')
-    # Character normalization
     text = text.replace("KH", "qa").replace("Z", "S")
     text = text.replace("SH", "SHa'a").replace("DH", "Dha'a")
     text = text.replace("ZamZam", "SamSam")
@@ -94,25 +89,35 @@ def normalize_text(text):
 def tts(text):
     paragraphs = text.strip().split("\n")
     audio_list = []
+    max_chars = 500  # Qiyaasta ugu badan 2 daqiiqo
+    warn_msg = ""
 
     for i, para in enumerate(paragraphs):
-        if not para.strip():
+        para = para.strip()
+        if not para:
             continue
-        norm_para = normalize_text(para)
-        inputs = tokenizer(norm_para, return_tensors="pt").to(device)
-        with torch.no_grad():
-            waveform = model(**inputs).waveform.squeeze().cpu().numpy()
 
-        # Add pause between paragraphs
-        if i < len(paragraphs) - 1:
+        if len(para) > max_chars:
+            warn_msg += f"❗ Qaybta {i+1} aad ayaa ka badan 2 daqiiqo. Waan kala jaray.\n"
+            sub_parts = [para[j:j+max_chars] for j in range(0, len(para), max_chars)]
+        else:
+            sub_parts = [para]
+
+        for part in sub_parts:
+            norm_para = normalize_text(part)
+            inputs = tokenizer(norm_para, return_tensors="pt").to(device)
+            with torch.no_grad():
+                waveform = model(**inputs).waveform.squeeze().cpu().numpy()
+
             pause = np.zeros(int(model.config.sampling_rate * 0.8))  # 0.8s pause
             audio_list.append(np.concatenate((waveform, pause)))
-        else:
-            audio_list.append(waveform)
 
     final_audio = np.concatenate(audio_list)
     filename = "output.wav"
     scipy.io.wavfile.write(filename, rate=model.config.sampling_rate, data=(final_audio * 32767).astype(np.int16))
+
+    if warn_msg:
+        print(warn_msg)
     return filename
 
 # Gradio interface
@@ -121,5 +126,5 @@ gr.Interface(
     inputs=gr.Textbox(label="Geli qoraal Soomaali ah", lines=10, placeholder="Ku qor 1 ama in ka badan paragraph..."),
     outputs=gr.Audio(label="Codka TTS"),
     title="Somali TTS",
-    description="Ku qor qoraal Soomaaliyeed si aad u maqasho cod dabiici ah."
+    description="Ku qor qoraal Soomaaliyeed si aad u maqasho cod dabiici ah. Qoraalka ha ka badnaan 2 daqiiqo per jumlad."
 ).launch()