Spaces:
Running
Running
| import torch | |
| import soundfile as sf | |
| import uuid | |
| import gradio as gr | |
| import numpy as np | |
| import re | |
| from parler_tts import ParlerTTSForConditionalGeneration | |
| from transformers import AutoTokenizer | |
| # Load model and tokenizers | |
| model_name = "ai4bharat/indic-parler-tts" | |
| device = "cpu" | |
| print("Loading model...") | |
| model = ParlerTTSForConditionalGeneration.from_pretrained(model_name).to(device).eval() | |
| tokenizer = AutoTokenizer.from_pretrained(model_name) | |
| desc_tokenizer = AutoTokenizer.from_pretrained(model.config.text_encoder._name_or_path) | |
| print("Applying dynamic quantization...") | |
| quantized_model = torch.quantization.quantize_dynamic( | |
| model, | |
| {torch.nn.Linear}, | |
| dtype=torch.qint8 | |
| ) | |
| # Sentence splitter | |
| def split_text(text, max_len=150): | |
| chunks = re.split(r'(?<=[.!?]) +', text) | |
| refined = [] | |
| for chunk in chunks: | |
| if len(chunk) <= max_len: | |
| refined.append(chunk) | |
| else: | |
| words = chunk.split() | |
| temp = [] | |
| buf_len = 0 | |
| for word in words: | |
| temp.append(word) | |
| buf_len += len(word) + 1 | |
| if buf_len > max_len: | |
| refined.append(' '.join(temp)) | |
| temp = [] | |
| buf_len = 0 | |
| if temp: | |
| refined.append(' '.join(temp)) | |
| return refined | |
| # Core TTS function | |
| def synthesize(language, text, gender, emotion, speed): | |
| description = ( | |
| f"A native {language.lower()} female speaker with an expressive tone." | |
| ) | |
| audio_chunks = [] | |
| text_chunks = split_text(text) | |
| for chunk in text_chunks: | |
| # New tokenization for each chunk | |
| desc_input = desc_tokenizer(description, return_tensors="pt").to(device) | |
| prompt_input = tokenizer(chunk, return_tensors="pt").to(device) | |
| with torch.no_grad(): | |
| output = quantized_model.generate( | |
| input_ids=desc_input.input_ids, | |
| attention_mask=desc_input.attention_mask, | |
| prompt_input_ids=prompt_input.input_ids, | |
| prompt_attention_mask=torch.ones_like(prompt_input.input_ids).to(device) | |
| ) | |
| audio = output.cpu().numpy().squeeze() | |
| audio_chunks.append(audio) | |
| full_audio = np.concatenate(audio_chunks) | |
| filename = f"{uuid.uuid4().hex}.wav" | |
| sf.write(filename, full_audio, quantized_model.config.sampling_rate) | |
| return filename | |
| # Gradio UI | |
| iface = gr.Interface( | |
| fn=synthesize, | |
| inputs=[ | |
| gr.Dropdown(["Malayalam", "Hindi", "Tamil", "English", "Kannada"], label="Language"), | |
| gr.Textbox(label="Text to Synthesize", lines=6, placeholder="Enter your sentence here..."), | |
| # gr.Radio(["Male", "Female"], label="Speaker Gender"), | |
| # gr.Dropdown(["Neutral", "Happy", "Sad", "Angry"], label="Emotion"), | |
| # gr.Dropdown(["Slow", "Moderate", "Fast"], label="Speaking Rate"), | |
| #gr.Dropdown(["Low", "Normal", "High"], label="Pitch"), | |
| #gr.Dropdown(["Basic", "Refined"], label="Voice Quality"), | |
| ], | |
| outputs=gr.Audio(type="filepath", label="Synthesized Speech"), | |
| title="Multilingual Indic TTS (Quantized + Chunked)", | |
| description="CPU-based TTS with quantized Parler-TTS and chunked input for Malayalam, Hindi, Tamil, and English.", | |
| ) | |
| iface.launch() |