Update app.py

app.py

@@ -1,30 +1,141 @@
 import gradio as gr
 import torch
-from transformers import SpeechT5ForTextToSpeech, SpeechT5Processor, SpeechT5HifiGan
 import soundfile as sf
+import spaces
+import os
+import numpy as np
+import re
+from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
+from speechbrain.pretrained import EncoderClassifier
+from datasets import load_dataset
 
+device = "cuda" if torch.cuda.is_available() else "cpu"
 
-model = SpeechT5ForTextToSpeech.from_pretrained("Beehzod/speecht5_finetuned_uz_customData2")
-processor = SpeechT5Processor.from_pretrained("Beehzod/speecht5_finetuned_uz_customData2")
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+def load_models_and_data():
+    model_name = "microsoft/speecht5_tts"
+    processor = SpeechT5Processor.from_pretrained(model_name)
+    model = SpeechT5ForTextToSpeech.from_pretrained("Beehzod/speecht5_finetuned_uz_customData2").to(device)
+    vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+    
+    spk_model_name = "speechbrain/spkrec-xvect-voxceleb"
+    speaker_model = EncoderClassifier.from_hparams(
+        source=spk_model_name,
+        run_opts={"device": device},
+        savedir=os.path.join("/tmp", spk_model_name),
+    )
+    
+    # Load a sample from a dataset for default embedding
+    dataset = load_dataset("Beehzod/UzTTS_data", split="train")
+    example = dataset[304]
+    
+    return model, processor, vocoder, speaker_model, example
 
-speaker_embeddings = torch.zeros((1, 512))
+model, processor, vocoder, speaker_model, default_example = load_models_and_data()
 
-def text_to_speech(text):
-  
-    inputs = processor(text=text, return_tensors="pt")
-    speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
-    output_path = "output.wav"
-    sf.write(output_path, speech.numpy(), 16000)
-    return output_path
+def create_speaker_embedding(waveform):
+    with torch.no_grad():
+        speaker_embeddings = speaker_model.encode_batch(torch.tensor(waveform).unsqueeze(0).to(device))
+        speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2)
+        speaker_embeddings = speaker_embeddings.squeeze()
+    return speaker_embeddings
 
+def prepare_default_embedding(example):
+    audio = example["audio"]
+    return create_speaker_embedding(audio["array"])
 
-interface = gr.Interface(
+default_embedding = prepare_default_embedding(default_example)
+
+replacements = [
+    ("а", "a"), ("б", "b"), ("в", "v"), ("г", "g"),
+    ("д", "d"), ("е", "e"), ("л", "l"), ("м", "m"),
+    ("о", "o"), ("р", "r"), ("с", "s"), ("т", "t"),
+    ("у", "u"), ("х", "x"), ("ю", "yu"), ("ё", "yo")
+]
+
+
+number_words = {
+    0: "nol", 1: "bir", 2: "ikki", 3: "uch", 4: "to'rt", 5: "besh", 6: "olti", 7: "yetti", 8: "sakkiz", 9: "to'qqiz",
+    10: "o'n", 11: "o'n bir", 12: "o'n ikki", 13: "o'n uch", 14: "o'n to'rt", 15: "o'n besh", 16: "o'n oltı", 17: "o'n yetti",
+    18: "o'n sakkiz", 19: "o'n toqqiz", 20: "yigirma", 30: "o'ttiz", 40: "qirq", 50: "ellik", 60: "oltmish", 70: "yetmish",
+    80: "sakson", 90: "to'qson", 100: "yuz", 1000: "ming", 1000000: "million"
+}
+def number_to_words(number):
+    if number < 20:
+        return number_words[number]
+    elif number < 100:
+        tens, unit = divmod(number, 10)
+        return number_words[tens * 10] + (" " + number_words[unit] if unit else "")
+    elif number < 1000:
+        hundreds, remainder = divmod(number, 100)
+        return (number_words[hundreds] + " yuz" if hundreds > 1 else "yuz") + (" " + number_to_words(remainder) if remainder else "")
+    elif number < 1000000:
+        thousands, remainder = divmod(number, 1000)
+        return (number_to_words(thousands) + " ming" if thousands > 1 else "ming") + (" " + number_to_words(remainder) if remainder else "")
+    elif number < 1000000000:
+        millions, remainder = divmod(number, 1000000)
+        return number_to_words(millions) + " million" + (" " + number_to_words(remainder) if remainder else "")
+    elif number < 1000000000000:
+        billions, remainder = divmod(number, 1000000000)
+        return number_to_words(billions) + " milliard" + (" " + number_to_words(remainder) if remainder else "")
+    else:
+        return str(number)
+
+
+def replace_numbers_with_words(text):
+    def replace(match):
+        number = int(match.group())
+        return number_to_words(number)
+
+    # Find the numbers and change with words.
+    result = re.sub(r'\b\d+\b', replace, text)
+
+    return result
+
+def normalize_text(text):
+    # Convert to lowercase
+    text = text.lower()
+    
+    # Replace numbers with words
+    text = replace_numbers_with_words(text)
+    
+    # Apply character replacements
+    for old, new in replacements:
+        text = text.replace(old, new)
+    
+    # Remove punctuation
+    text = re.sub(r'[^\w\s]', '', text)
+    
+    return text
+
+@spaces.GPU(duration=60)
+def text_to_speech(text, audio_file=None):
+    # Normalize the input text
+    normalized_text = normalize_text(text)
+    
+    # Prepare the input for the model
+    inputs = processor(text=normalized_text, return_tensors="pt").to(device)
+    
+    # Use the default speaker embedding
+    speaker_embeddings = default_embedding
+    
+    # Generate speech
+    with torch.no_grad():
+        speech = model.generate_speech(inputs["input_ids"], speaker_embeddings.unsqueeze(0), vocoder=vocoder)
+    
+    speech_np = speech.cpu().numpy()
+
+    return (16000, speech_np)
+    
+iface = gr.Interface(
     fn=text_to_speech,
-    inputs="text",
-    outputs="audio",
-    title="Uzbek Text-to-Speech Generator",
-    description="Enter Uzbek text and generate speech using the finetuned SpeechT5 model."
+    inputs=[
+        gr.Textbox(label="Enter Uzbek text to convert to speech")
+    ],
+    outputs=[
+        gr.Audio(label="Generated Speech", type="numpy")
+    ],
+    title="Uzbek SpeechT5 Text-to-Speech Demo",
+    description="Enter Uzbek text, and listen to the generated speech."
 )
 
-interface.launch()
+iface.launch(share=True)