Update README.md and create Gradio app for testing

README.md

@@ -1,8 +1,8 @@
 ---
-title: Svara Tts
-emoji: 👁
-colorFrom: pink
-colorTo: pink
+title: Svara TTS
+emoji: 🗣️
+colorFrom: blue
+colorTo: violet
 sdk: gradio
 sdk_version: 5.49.1
 app_file: app.py
@@ -10,4 +10,45 @@ pinned: false
 license: apache-2.0
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🗣️ Svara-TTS
+
+**An open multilingual text-to-speech (TTS) model bringing expressive, human-like speech to India’s languages.**
+
+Svara-TTS is trained on **1,900+ hours** of diverse Indian speech across **17 languages**, with balanced male and female voices.  
+It captures **emotion**, **tone**, and **rhythm**, producing natural, human-like speech for real-world Indian use cases.
+
+> 💡 Supports `<happy>`, `<sad>`, `<fear>`, `<relief>`, and other emotion tags — plus multilingual and zero-shot voice cloning.
+
+---
+
+## 🌍 Supported Languages
+
+Hindi, Marathi, Tamil, Telugu, Bengali, Kannada, Gujarati, Malayalam, Punjabi, Assamese, Odia, Bhojpuri, Chhattisgarhi, Maithili, Magahi, Nepali, and Indian English.
+
+---
+
+## 🎧 Try These
+
+Hindi (Female): अरे वाह! आज तो मौसम बहुत ही सुहावना लग रहा है। 
+Tamil (Male): இன்று அலுவலகத்தில் பெரிய கூட்டம் இருந்தது, ஆனா எல்லாம் நன்றாக முடிந்தது. 
+Telugu (Female): నిజం చెప్పాలంటే, ఈ రోజు కొంచెం టెన్షన్ గా ఉంది… కానీ అన్ని బాగానే జరుగుతాయని నమ్మకం ఉంది. 
+English (Female): Sometimes it’s not about being perfect… it’s about being real. 
+
+---
+
+### ⚙️ Technical Overview
+
+- Model: **kenpath/svara-tts-v1**  
+- Codec: [SNAC 24kHz](https://huggingface.co/hubertsiuzdak/snac_24khz)  
+- Sampling rate: **24,000 Hz**  
+- Framework: **PyTorch + Transformers**
+
+---
+
+### 🧩 About
+
+Built by [**Kenpath Technologies**](https://kenpath.ai)  
+Trained on open datasets including SYSPIN, SPICOR, RASA, and IndicTTS  
+Released under **Apache-2.0 License**
+
+> “Svara is how India sounds — open, expressive, and human.”

app.py

@@ -0,0 +1,182 @@
+import spaces
+from snac import SNAC
+import torch
+import gradio as gr
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from huggingface_hub import snapshot_download
+from dotenv import load_dotenv
+load_dotenv()
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+print("Loading SNAC model...")
+snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to(device)
+
+model_name = "kenpath/svara-tts-v1"
+
+snapshot_download(
+    repo_id=model_name,
+    allow_patterns=["config.json", "*.safetensors", "model.safetensors.index.json"],
+    ignore_patterns=["optimizer.pt", "pytorch_model.bin", "training_args.bin", "scheduler.pt"]
+)
+
+model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16).to(device)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+print(f"Svara model loaded to {device}")
+
+# --------------------------
+# Language and Gender Setup
+# --------------------------
+LANGUAGES = {
+    "Assamese (অসমীয়া)": "Assamese",
+    "Bengali (বাংলা)": "Bengali",
+    "Bhojpuri (भोजपुरी)": "Bhojpuri",
+    "Chhattisgarhi (छत्तीसगढ़ी)": "Chhattisgarhi",
+    "Gujarati (ગુજરાતી)": "Gujarati",
+    "Hindi (हिन्दी)": "Hindi",
+    "Kannada (ಕನ್ನಡ)": "Kannada",
+    "Maithili (मैथिली)": "Maithili",
+    "Magahi (मगही)": "Magahi",
+    "Malayalam (മലയാളം)": "Malayalam",
+    "Marathi (मराठी)": "Marathi",
+    "Nepali (नेपाली)": "Nepali",
+    "Odia (ଓଡ଼ିଆ)": "Odia",
+    "Punjabi (ਪੰਜਾਬੀ)": "Punjabi",
+    "Tamil (தமிழ்)": "Tamil",
+    "Telugu (తెలుగు)": "Telugu",
+    "English (Indian)": "English"
+}
+GENDERS = ["Male", "Female"]
+
+# Emotion tags for user help text
+EMOTIVE_TAGS = ["<happy>", "<sad>", "<fear>", "<surprise>", "<calm>", "<angry>", "<relief>", "<hopeful>", "<thoughtful>"]
+
+# --------------------------
+# Prompt Preparation
+# --------------------------
+def process_prompt(language, gender, text, tokenizer, device):
+    lang_label = LANGUAGES.get(language, "English")
+    prompt = f"{lang_label} ({gender}): {text}"
+    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+    start_token = torch.tensor([[128259]], dtype=torch.int64)
+    end_tokens = torch.tensor([[128009, 128260]], dtype=torch.int64)
+    modified_input_ids = torch.cat([start_token, input_ids, end_tokens], dim=1)
+    attention_mask = torch.ones_like(modified_input_ids)
+    return modified_input_ids.to(device), attention_mask.to(device)
+
+# --------------------------
+# Generation Functions
+# --------------------------
+def parse_output(generated_ids):
+    token_to_find, token_to_remove = 128257, 128258
+    token_indices = (generated_ids == token_to_find).nonzero(as_tuple=True)
+    cropped_tensor = generated_ids[:, token_indices[1][-1] + 1:] if len(token_indices[1]) > 0 else generated_ids
+    processed_rows = [row[row != token_to_remove] for row in cropped_tensor]
+    row = processed_rows[0]
+    trimmed_row = row[: (row.size(0) // 7) * 7]
+    trimmed_row = [t - 128266 for t in trimmed_row]
+    return trimmed_row
+
+def redistribute_codes(code_list, snac_model):
+    layer_1, layer_2, layer_3 = [], [], []
+    for i in range((len(code_list) + 1) // 7):
+        layer_1.append(code_list[7*i])
+        layer_2.append(code_list[7*i+1]-4096)
+        layer_3.append(code_list[7*i+2]-(2*4096))
+        layer_3.append(code_list[7*i+3]-(3*4096))
+        layer_2.append(code_list[7*i+4]-(4*4096))
+        layer_3.append(code_list[7*i+5]-(5*4096))
+        layer_3.append(code_list[7*i+6]-(6*4096))
+    codes = [torch.tensor(x, device=device).unsqueeze(0) for x in [layer_1, layer_2, layer_3]]
+    return snac_model.decode(codes).detach().squeeze().cpu().numpy()
+
+@spaces.GPU()
+def generate_speech(language, gender, text, temperature, top_p, repetition_penalty, max_new_tokens, progress=gr.Progress()):
+    if not text.strip():
+        return None
+    try:
+        progress(0.1, "Processing text...")
+        input_ids, attention_mask = process_prompt(language, gender, text, tokenizer, device)
+        progress(0.3, "Generating speech tokens...")
+        with torch.no_grad():
+            generated_ids = model.generate(
+                input_ids=input_ids, attention_mask=attention_mask,
+                max_new_tokens=max_new_tokens, do_sample=True,
+                temperature=temperature, top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                num_return_sequences=1, eos_token_id=128258
+            )
+        progress(0.6, "Parsing output...")
+        code_list = parse_output(generated_ids)
+        progress(0.8, "Converting to audio...")
+        audio_samples = redistribute_codes(code_list, snac_model)
+        return (24000, audio_samples)
+    except Exception as e:
+        print(f"Error generating speech: {e}")
+        return None
+
+# --------------------------
+# Example Prompts
+# --------------------------
+examples = [
+    ["Hindi (हिन्दी)", "Female", "अरे वाह! आज तो मौसम बहुत ही सुहावना लग रहा है। <happy>", 0.6, 0.95, 1.1, 1200],
+    ["Marathi (मराठी)", "Male", "खरंच सांगतो, आजचा दिवस खूप छान गेला! <happy>", 0.6, 0.95, 1.1, 1200],
+    ["Tamil (தமிழ்)", "Male", "இன்று அலுவலகத்தில் பெரிய கூட்டம் இருந்தது, ஆனா எல்லாம் நன்றாக முடிந்தது. <relief>", 0.65, 0.95, 1.1, 1200],
+    ["Bengali (বাংলা)", "Female", "আজ অফিসে এত কাজ ছিল যে মাথা ধরেছে! একটু বিশ্রাম নেব ভাবছি। <sad>", 0.7, 0.9, 1.1, 1200],
+    ["Telugu (తెలుగు)", "Female", "నిజం చెప్పాలంటే, ఈ రోజు కొంచెం టెన్షన్ గా ఉంది... కానీ అన్ని బాగానే జరుగుతాయని నమ్మకం ఉంది. <fear>", 0.7, 0.95, 1.1, 1200],
+    ["Gujarati (ગુજરાતી)", "Female", "અરે વાહ, આજે તો પૂરો દિવસ ફક્ત વરસાદ જ પડ્યો! ચા પી લઈએ ને? <joy>", 0.6, 0.95, 1.1, 1200],
+    ["Kannada (ಕನ್ನಡ)", "Male", "ಹೌದು, ನಿನ್ನೆ ರಾತ್ರಿ ತುಂಬಾ ಮಳೆ ಬಿತ್ತು. ಬೆಳಿಗ್ಗೆ ರಸ್ತೆಗಳಲ್ಲಿ ನೀರು ತುಂಬಿದೆ. <surprise>", 0.65, 0.95, 1.1, 1200],
+    ["English (Indian)", "Female", "Sometimes it's not about being perfect... it's about being real, and just a little bit messy. <reflective>", 0.6, 0.95, 1.1, 1200],
+]
+
+# --------------------------
+# Gradio UI
+# --------------------------
+with gr.Blocks(title="Svara Multilingual TTS") as demo:
+    gr.Markdown(f"""
+    # 🎵 Svara-TTS  
+    *An open multilingual TTS model for expressive, human-like speech across India's languages.*
+
+    **Tips:**
+    - Add emotion tags like {", ".join(EMOTIVE_TAGS)} for expressive speech.  
+    - Use longer, natural sentences for better prosody.  
+    - Choose a language and gender before generating.
+    """)
+
+    with gr.Row():
+        with gr.Column(scale=3):
+            lang = gr.Dropdown(choices=list(LANGUAGES.keys()), value="Hindi (हिन्दी)", label="Language")
+            gender = gr.Dropdown(choices=GENDERS, value="Female", label="Gender")
+            text_input = gr.Textbox(label="Text to speak", placeholder="Type your text...", lines=5)
+
+            with gr.Accordion("Advanced Settings", open=False):
+                temperature = gr.Slider(0.1, 1.5, 0.6, 0.05, label="Temperature")
+                top_p = gr.Slider(0.1, 1.0, 0.95, 0.05, label="Top-p")
+                repetition_penalty = gr.Slider(1.0, 2.0, 1.1, 0.05, label="Repetition Penalty")
+                max_new_tokens = gr.Slider(100, 2000, 1200, 100, label="Max Tokens")
+
+            with gr.Row():
+                submit = gr.Button("Generate Speech", variant="primary")
+                clear = gr.Button("Clear")
+
+        with gr.Column(scale=2):
+            audio_output = gr.Audio(label="Generated Speech", type="numpy")
+
+    gr.Examples(
+        examples=examples,
+        inputs=[lang, gender, text_input, temperature, top_p, repetition_penalty, max_new_tokens],
+        outputs=audio_output,
+        fn=generate_speech,
+        cache_examples=True,
+    )
+
+    submit.click(
+        fn=generate_speech,
+        inputs=[lang, gender, text_input, temperature, top_p, repetition_penalty, max_new_tokens],
+        outputs=audio_output,
+    )
+
+    clear.click(fn=lambda: (None, None), inputs=[], outputs=[text_input, audio_output])
+
+if __name__ == "__main__":
+    demo.queue().launch(share=False, ssr_mode=False)

requirements.txt

@@ -0,0 +1,5 @@
+snac
+python-dotenv
+transformers
+torch
+spaces