Update app.py

app.py

@@ -1,46 +1,94 @@
-from fastapi import FastAPI
-from pydantic import BaseModel
-from transformers import VitsModel, AutoTokenizer
+import gradio as gr
 import torch
-import base64
-import io
-import soundfile as sf
-
-app = FastAPI()
-
-# Load model & tokenizer hal mar marka server-ka bilaabmo
-model_name = "facebook/mms-tts-som"
-model = VitsModel.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name)
+import numpy as np
+import scipy.io.wavfile
+from transformers import VitsModel, AutoTokenizer
+import re
 
+# Load fine-tuned model from Hugging Face Hub or local path
+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()
 
-class TTSRequest(BaseModel):
-    text: str
+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"
+}
 
-@app.post("/tts")
-async def tts(request: TTSRequest):
-    text = request.text
-    # Tokenize input text
-    tokens = tokenizer(text, return_tensors="pt")
-    tokens = {k: v.to(device) for k, v in tokens.items()}
+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)
 
-    # Generate speech waveform tensor
-    with torch.no_grad():
-        audio = model.generate_speech(tokens['input_ids'])
-
-    # Convert tensor to numpy array
-    audio_np = audio.squeeze().cpu().numpy()
+def normalize_text(text):
+    numbers = re.findall(r'\d+', text)
+    for num in numbers:
+        text = text.replace(num, number_to_words(num))
+    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
 
-    # Write to WAV buffer
-    buffer = io.BytesIO()
-    sf.write(buffer, audio_np, samplerate=22050, format='WAV')
-    wav_bytes = buffer.getvalue()
-
-    # Encode wav bytes to base64
-    b64_audio = base64.b64encode(wav_bytes).decode('utf-8')
+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()
+    filename = "output.wav"
+    scipy.io.wavfile.write(filename, rate=model.config.sampling_rate, data=(waveform * 32767).astype(np.int16))
+    return filename
 
-    # Return base64 audio string with wav header
-    return {"audio": f"data:audio/wav;base64,{b64_audio}"}
+gr.Interface(
+    fn=tts,
+    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.",
+).launch()
+