added files

app.py

@@ -0,0 +1,101 @@
+import os
+import torch
+import time
+from parler_tts import ParlerTTSForConditionalGeneration
+from transformers import AutoTokenizer
+import soundfile as sf
+from google.generativeai import GenerativeModel, configure
+import gradio as gr
+
+# Initialize with prints
+print("⚡ Initializing models...")
+start_load = time.time()
+
+# 1. Load Gemini
+GEMINI_KEY = os.environ.get('GEMINI_API_KEY')
+configure(api_key=GEMINI_KEY)
+gemini = GenerativeModel('gemini-2.0-flash')
+print(f"   ✅ Gemini loaded (device: {'GPU' if torch.cuda.is_available() else 'CPU'})")
+
+# 2. Load Indic-TTS
+device = "cuda" if torch.cuda.is_available() else "cpu"
+tts_model = ParlerTTSForConditionalGeneration.from_pretrained("ai4bharat/indic-parler-tts").to(device)
+tts_tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-parler-tts")
+desc_tokenizer = AutoTokenizer.from_pretrained(tts_model.config.text_encoder._name_or_path)
+print(f"   ✅ Indic-TTS loaded in {time.time() - start_load:.2f}s\n")
+
+def hinglish_to_devnagri(text):
+    try:
+        print(f"🔠 Converting Hinglish to Devnagri: '{text[:30]}...'")
+        start = time.time()
+        
+        response = gemini.generate_content(
+            f"""
+            Analyze this mixed-language text containing English and a regional language:
+            "{text}"
+
+            Perform:
+            1. Detect the regional language (e.g. Hindi, Tamil, Bengali) if you don't find regional language, convert to Hindi (pure devanagari script).
+            2. Convert to pure regional language script (Devanagari/Tamil/Bangla)
+            3. Preserve complex words (technical/medical terms) in their original form
+            4. Maintain natural flow and meaning
+            5. Remove the code if you find them in backticks ```.
+
+            Rules:
+            - Keep proper nouns unchanged
+            - Use colloquial spellings (e.g. "कॉलेज" not "विद्यालय" for "college")
+
+            Output ONLY the converted text in the detected script.
+            """
+        )
+        
+        print(f"   ✓ Translation done in {time.time() - start:.2f}s")
+        return response.text
+    except Exception as e:
+        print(f"❌ Gemini error: {str(e)}")
+        raise gr.Error(f"Gemini error: {str(e)}")
+
+def generate_speech(text):
+    print("\n" + "="*50)
+    print("🎤 Starting Hinglish-to-Speech pipeline")
+    
+    # 1. Text Conversion
+    hindi_text = hinglish_to_devnagri(text)
+    print(f"   Hindi text: {hindi_text[:50]}...")
+    
+    # 2. Audio Generation
+    print("\n🔊 Generating audio...")
+    start_audio = time.time()
+    
+    desc = "एक महिला वक्ता स्पष्ट हिंदी में बोल रही हैं"
+    desc_inputs = desc_tokenizer(desc, return_tensors="pt").to(device)
+    text_inputs = tts_tokenizer(hindi_text, return_tensors="pt").to(device)
+    
+    audio = tts_model.generate(
+        input_ids=desc_inputs.input_ids,
+        attention_mask=desc_inputs.attention_mask,
+        prompt_input_ids=text_inputs.input_ids,
+        prompt_attention_mask=text_inputs.attention_mask
+    )
+    
+    # 3. Save Output
+    sf.write("output.wav", audio.cpu().numpy().squeeze(), tts_model.config.sampling_rate)
+    print(f"\n💾 Audio generated in {time.time() - start_audio:.2f}s")
+    print("="*50 + "\n")
+    
+    return "output.wav", hindi_text
+
+# Gradio UI
+with gr.Blocks() as app:
+    gr.Markdown("## 🚀 Hinglish-to-Speech (Gemini + Indic-TTS)")
+    with gr.Row():
+        inp = gr.Textbox(label="Enter Hinglish Text", placeholder="Aaj mood nahi hai...")
+        btn = gr.Button("Generate")
+    with gr.Row():
+        audio_out = gr.Audio(label="Speech Output")
+        text_out = gr.Textbox(label="Devnagri Translation")
+    
+    btn.click(fn=generate_speech, inputs=inp, outputs=[audio_out, text_out])
+
+print("\n🚀 App ready! Waiting for input...")
+app.launch()

requirements.txt

@@ -0,0 +1,8 @@
+torch>=2.0.0
+transformers>=4.30.0
+parler-tts>=0.1.0
+soundfile>=0.12.0
+google-generativeai>=0.3.0
+gradio>=3.40.0
+
+