Upload app.py

app.py

@@ -0,0 +1,224 @@
+# -*- coding: utf-8 -*-
+"""app.ipynb
+
+Automatically generated by Colab.
+
+Original file is located at
+    https://colab.research.google.com/drive/1E-6YoHEsjoKc6FVXzQHbtnRja-68brYc
+
+numpy==1.23.5
+transformers
+datasets
+soundfile
+torch
+torchaudio
+sentencepiece
+speechbrain==0.5.16
+librosa
+"""
+
+!pip install transformers datasets soundfile torch torchaudio sentencepiece speechbrain librosa
+
+!pip install -q gradio
+
+!pip install -q spaces
+
+import gradio as gr
+import torch
+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"
+
+def load_models_and_data():
+    model_name = "microsoft/speecht5_tts"
+    processor = SpeechT5Processor.from_pretrained(model_name)
+    model = SpeechT5ForTextToSpeech.from_pretrained("Moustapha91/speecht5_tts__v3_sn").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("Alwaly/Wolof_TTS", split="train")
+    dataset = dataset.select(range(12003))
+    dataset = dataset.train_test_split(test_size=0.1)
+    example = dataset['test'][8]
+
+    return model, processor, vocoder, speaker_model, example
+
+model, processor, vocoder, speaker_model, default_example = load_models_and_data()
+
+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"])
+
+default_embedding = prepare_default_embedding(default_example)
+
+replacements = [
+    ("а", "a"),
+    ("à", "a"),
+    ('á', "a"),
+    ('ã', "a"),
+    ('ä', "a"),
+    ('α', "a"),
+    ("ă", "a"),
+    ("â", "a"),
+    ('ā', "a"),
+    ('ą', "a"),
+    ('ɓ', "b"),
+    ('β', "b"),
+    ("ç", "c"),
+    ('с', "c"),
+    ('ɗ', "d"),
+    ('đ', "d"),
+    ("è", "e"),
+    ('ẽ', "e"),
+    ("ë", "e"),
+    ("ð", "e"),
+    ('έ', "e"),
+    ('ɐ', "e"),
+    ('ə', "e"),
+    ('ξ', "e"),
+    ('н', "h"),
+    ("î", "i"),
+    ('í', "i"),
+    ("ị", "i"),
+    ("ï", "i"),
+    ('к', "k"),
+    ('ł', "l"),
+    ("ŋ", "n"),
+    ('ń', "n"),
+    ("ñ", "n"),
+    ('ῆ', "n"),
+    ('й', "n"),
+    ('η', "n"),
+    ("ó", "o"),
+    ('ồ', "o"),
+    ('ớ', "o"),
+    ('ὀ', "o"),
+    ("ô", "o"),
+    ("õ", "o"),
+    ('ò', "o"),
+    ('ø', "o"),
+    ('σ', "o"),
+    ("ο", "o"),
+    ("ο", "o"),
+    ('р', "p"),
+    ('ρ', "p"),
+    ("т", "t"),
+    ('ƭ', "t"),
+    ('ц', "u"),
+    ("ù", "u"),
+    ("û", "u"),
+    ('μ', "u"),
+    ('ш', "w"),
+#     ('у', "y"),
+#     ('ý', "y"),
+#     ('γ', "y"),
+]
+
+number_words = {
+    0: "dara", 1: "benn", 2: "ñaar", 3: "ñett", 4: "ñent", 5: "juróom ", 6: "juróom ak benn", 7: "juróom ak ñaar", 8: "juróom ak ñett", 9: "juróom ak ñent",
+    10: "fukk", 11: "fukk ak benn", 12: "fukk ak ñaar", 13: "fukk ak ñett", 14: "fukk ak ñent", 15: "fukk", 16: "fukk ak juróom ben", 17: "fukk ak juróom ñaar",
+    18: "fukk ak juróom ñett", 19: "fukk ak juróom ñent", 20: "ñaar fukk", 30: "ñett fukk", 40: "ñent fukk", 50: "juróom fukk", 60: "juróom benn fukk", 70: "juróom ñaar fukk",
+    80: "juróom ñett fukk", 90: "juróom ñent fukk", 100: "téeméer", 1000: " junni"
+}
+
+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] + " yüz" if hundreds > 1 else "yüz") + (" " + number_to_words(remainder) if remainder else "")
+    elif number < 1000000:
+        thousands, remainder = divmod(number, 1000)
+        return (number_to_words(thousands) + " bin" if thousands > 1 else "bin") + (" " + number_to_words(remainder) if remainder else "")
+    elif number < 1000000000:
+        millions, remainder = divmod(number, 1000000)
+        return number_to_words(millions) + " milyon" + (" " + number_to_words(remainder) if remainder else "")
+    elif number < 1000000000000:
+        billions, remainder = divmod(number, 1000000000)
+        return number_to_words(billions) + " milyar" + (" " + 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=[
+        gr.Textbox(label="Enter Woloftext to convert to speech")
+    ],
+    outputs=[
+        gr.Audio(label="Generated Speech", type="numpy")
+    ],
+    title="Wolof SpeechT5 Text-to-Speech Demo",
+    description="Enter Wolof text, and listen to the generated speech."
+)
+
+iface.launch(share=True)
+