Update app.py

app.py

@@ -1,3 +1,6 @@
+# WARNING: THIS CODE IS FOR ILLUSTRATION ONLY AND WILL NOT WORK.
+# The 'Somalitts/somali_tts_model' does not support voice cloning.
+
 import gradio as gr
 import torch
 import numpy as np
@@ -5,128 +8,66 @@ import scipy.io.wavfile
 from transformers import VitsModel, AutoTokenizer
 import re
 
-# Load model and tokenizer
+# --- The problem starts here ---
+# This model is a single-speaker model. It CANNOT clone voices.
+# To make this work, you would need a different model designed for cloning.
 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",
-    11: "toban iyo koow", 12: "toban iyo labo", 13: "toban iyo seddex",
-    14: "toban iyo afar", 15: "toban iyo shan", 16: "toban iyo lix",
-    17: "toban iyo todobo", 18: "toban iyo sideed", 19: "toban iyo sagaal",
-    20: "labaatan", 30: "sodon", 40: "afartan", 50: "konton",
-    60: "lixdan", 70: "todobaatan", 80: "sideetan", 90: "sagaashan",
-    100: "boqol", 1000: "kun"
-}
+# For this to work, you would need to upload your voice files to your Space
+# and provide the path here.
+YOUR_VOICE_SAMPLE_PATH = ["46.wav", "90.wav", "150.wav", "355.wav"]
 
-def number_to_words(number):
-    number = int(number)
-    if number < 20:
-        return number_words[number]
-    elif number < 100:
-        tens, unit = divmod(number, 10)
-        return number_words[tens * 10] + (" iyo " + number_words[unit] if unit else "")
-    elif number < 1000:
-        hundreds, remainder = divmod(number, 100)
-        part = (number_words[hundreds] + " boqol") if hundreds > 1 else "boqol"
-        if remainder:
-            part += " iyo " + number_to_words(remainder)
-        return part
-    elif number < 1000000:
-        thousands, remainder = divmod(number, 1000)
-        words = []
-        if thousands == 1:
-            words.append("kun")
-        else:
-            words.append(number_to_words(thousands) + " kun")
-        if remainder >= 100:
-            hundreds, rem2 = divmod(remainder, 100)
-            if hundreds:
-                boqol_text = (number_words[hundreds] + " boqol") if hundreds > 1 else "boqol"
-                words.append(boqol_text)
-            if rem2:
-                words.append("iyo " + number_to_words(rem2))
-        elif remainder:
-            words.append("iyo " + number_to_words(remainder))
-        return " ".join(words)
-    elif number < 1000000000:
-        millions, remainder = divmod(number, 1000000)
-        words = []
-        if millions == 1:
-            words.append("milyan")
-        else:
-            words.append(number_to_words(millions) + " milyan")
-        if remainder:
-            words.append(number_to_words(remainder))
-        return " ".join(words)
-    else:
-        return str(number)
-
-def normalize_text(text):
-    text = re.sub(r'(\d{1,3})(,\d{3})+', lambda m: m.group(0).replace(",", ""), text)
-    text = re.sub(r'\.\d+', '', text)
-    def replace_num(match):
-        return number_to_words(match.group())
-    text = re.sub(r'\d+', replace_num, text)
-    symbol_map = {
-        '$': 'doolar',
-        '=': 'egwal',
-        '+': 'balaas',
-        '#': 'haash'
-    }
-    for sym, word in symbol_map.items():
-        text = text.replace(sym, ' ' + word + ' ')
-    text = text.replace("KH", "qa").replace("Z", "S")
-    text = text.replace("SH", "SHa'a").replace("DH", "Dha'a")
-    text = text.replace("ZamZam", "samsam")
-    text = text.replace("Zamzam", "samsam")
-    text = text.replace("zamzam", "samsam")
-    return text
+# [Your other functions like number_to_words and normalize_text would remain here]
+# ...
 
 def tts(text):
+    # --- The core logic would need to change entirely ---
+    
+    # 1. THIS IS THE MISSING STEP:
+    # A real voice cloning model would need to extract voice characteristics
+    # from your audio file. The VitsModel you are using has NO such function.
+    #
+    # PSEUDO-CODE (DOES NOT EXIST FOR THIS MODEL):
+    # voice_characteristics = model.extract_speaker_embedding(YOUR_VOICE_SAMPLE_PATH)
+
     paragraphs = text.strip().split("\n")
     audio_list = []
-    max_chars = 500  # Qiyaasta ugu badan 2 daqiiqo
-    warn_msg = ""
+    
+    for para in paragraphs:
+        # [Text processing would be the same]
+        # ...
+        norm_para = normalize_text(para)
+        inputs = tokenizer(norm_para, return_tensors="pt").to(device)
 
-    for i, para in enumerate(paragraphs):
-        para = para.strip()
-        if not para:
-            continue
+        with torch.no_grad():
+            # 2. THIS IS THE SECOND MISSING STEP:
+            # You would need to pass your voice characteristics to the model.
+            # The current model does not accept a 'speaker_embedding' or similar argument.
+            #
+            # PSEUDO-CODE (DOES NOT EXIST FOR THIS MODEL):
+            # waveform = model(**inputs, speaker_embedding=voice_characteristics).waveform
+            
+            # The actual line of code below does not and cannot use your voice:
+            waveform = model(**inputs).waveform.squeeze().cpu().numpy()
 
-        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)))
+        pause = np.zeros(int(model.config.sampling_rate * 0.8))
+        audio_list.append(np.concatenate((waveform, pause)))
 
     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
+# The interface would also need an input for the audio file.
 gr.Interface(
     fn=tts,
-    inputs=gr.Textbox(label="Geli qoraal Soomaali ah", lines=10, placeholder="Ku qor 1 ama in ka badan paragraph..."),
+    inputs=gr.Textbox(label="Geli qoraal Soomaali ah"),
     outputs=gr.Audio(label="Codka TTS"),
-    title="Somali TTS",
-    description="Ku qor qoraal Soomaaliyeed si aad u maqasho cod dabiici ah. Qoraalka ha ka badnaan 2 daqiiqo per jumlad."
+    title="Somali TTS (Non-Cloning)"
 ).launch()