Update app.py

app.py

@@ -5,13 +5,14 @@ import scipy.io.wavfile
 from transformers import VitsModel, AutoTokenizer
 import re
 
-# Load fine-tuned model from Hugging Face Hub or local path
+# Load model and tokenizer
 model = VitsModel.from_pretrained("Somali-tts/somali_tts_model")
 tokenizer = AutoTokenizer.from_pretrained("saleolow/somali-mms-tts")
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 model.to(device)
 model.eval()
 
+# Numbers in Somali
 number_words = {
     0: "eber", 1: "koow", 2: "labo", 3: "seddex", 4: "afar", 5: "shan",
     6: "lix", 7: "todobo", 8: "sideed", 9: "sagaal", 10: "toban",
@@ -67,28 +68,58 @@ def number_to_words(number):
         return str(number)
 
 def normalize_text(text):
-    numbers = re.findall(r'\d+', text)
-    for num in numbers:
-        text = text.replace(num, number_to_words(num))
+    # 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',
+        '+': 'balaas',
+        '#': 'haash'
+    }
+    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")
     return text
 
 def tts(text):
-    text = normalize_text(text)
-    inputs = tokenizer(text, return_tensors="pt").to(device)
-    with torch.no_grad():
-        waveform = model(**inputs).waveform.squeeze().cpu().numpy()
+    paragraphs = text.strip().split("\n")
+    audio_list = []
+
+    for i, para in enumerate(paragraphs):
+        if not para.strip():
+            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:
+            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=(waveform * 32767).astype(np.int16))
+    scipy.io.wavfile.write(filename, rate=model.config.sampling_rate, data=(final_audio * 32767).astype(np.int16))
     return filename
 
+# Gradio interface
 gr.Interface(
     fn=tts,
-    inputs=gr.Textbox(label="Geli qoraal Soomaali ah"),
+    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."
 ).launch()
-