Update app.py
Browse files
app.py
CHANGED
|
@@ -1,49 +1,57 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
| 3 |
from transformers import VitsModel, AutoTokenizer
|
| 4 |
-
import scipy.io.wavfile as wav
|
| 5 |
-
import numpy as np
|
| 6 |
-
import tempfile
|
| 7 |
|
| 8 |
-
#
|
| 9 |
-
|
|
|
|
| 10 |
model = VitsModel.from_pretrained(model_name)
|
| 11 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 12 |
|
| 13 |
-
#
|
| 14 |
-
def
|
| 15 |
-
#
|
| 16 |
-
|
| 17 |
-
return "براہِ مہربانی، درست اردو متن درج کریں۔"
|
| 18 |
-
|
| 19 |
-
# ٹوکنائزیشن
|
| 20 |
-
inputs = tokenizer(urdu_text, return_tensors="pt", padding=True, truncation=True, max_length=512)
|
| 21 |
-
|
| 22 |
-
# ٹوکنائزیشن کے بعد ان پٹ کی جانچ
|
| 23 |
-
if inputs["input_ids"].size(1) == 0:
|
| 24 |
-
return "ٹوکنائزیشن کے بعد ان پٹ خالی ہے۔ براہِ مہربانی، متن کی جانچ کریں اور دوبارہ کوشش کریں۔"
|
| 25 |
-
|
| 26 |
-
# input_ids کو LongTensor میں تبدیل کریں
|
| 27 |
-
inputs["input_ids"] = inputs["input_ids"].to(torch.long)
|
| 28 |
|
|
|
|
| 29 |
with torch.no_grad():
|
| 30 |
-
output = model(**inputs).waveform
|
| 31 |
-
|
| 32 |
-
#
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
| 3 |
from transformers import VitsModel, AutoTokenizer
|
|
|
|
|
|
|
|
|
|
| 4 |
|
| 5 |
+
# 1. Load the model (Nastaliq-based) and tokenizer
|
| 6 |
+
# This checkpoint is intended for Urdu text in its traditional (Nastaliq) script.
|
| 7 |
+
model_name = "facebook/mms-tts-urd"
|
| 8 |
model = VitsModel.from_pretrained(model_name)
|
| 9 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 10 |
|
| 11 |
+
# 2. Define the inference function
|
| 12 |
+
def generate_urdu_speech(urdu_text):
|
| 13 |
+
# Tokenize the input text
|
| 14 |
+
inputs = tokenizer(urdu_text, return_tensors="pt")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
+
# Perform inference with the model
|
| 17 |
with torch.no_grad():
|
| 18 |
+
output = model(**inputs).waveform
|
| 19 |
+
|
| 20 |
+
# Convert PyTorch tensor to NumPy array
|
| 21 |
+
waveform = output.squeeze().cpu().numpy()
|
| 22 |
+
sample_rate = model.config.sampling_rate
|
| 23 |
+
|
| 24 |
+
# Gradio’s Audio component expects (sample_rate, audio_data)
|
| 25 |
+
return (sample_rate, waveform)
|
| 26 |
+
|
| 27 |
+
# 3. Build the Gradio interface
|
| 28 |
+
with gr.Blocks() as demo:
|
| 29 |
+
gr.Markdown("""
|
| 30 |
+
# Urdu TTS Demo (Nastaliq Script)
|
| 31 |
+
|
| 32 |
+
Enter text in Urdu (Nastaliq) script, and this demo will synthesize speech using the Facebook MMS TTS model for Urdu.
|
| 33 |
+
""")
|
| 34 |
+
|
| 35 |
+
# Text input for Urdu (Nastaliq)
|
| 36 |
+
text_input = gr.Textbox(
|
| 37 |
+
label="Enter Urdu text",
|
| 38 |
+
placeholder="مثال کے طور پر...",
|
| 39 |
+
lines=3
|
| 40 |
+
)
|
| 41 |
+
|
| 42 |
+
# Audio output
|
| 43 |
+
audio_output = gr.Audio(label="Generated Urdu Speech", type="numpy")
|
| 44 |
+
|
| 45 |
+
# Generate button
|
| 46 |
+
generate_button = gr.Button("Generate Speech")
|
| 47 |
+
|
| 48 |
+
# Wire up the button to the function
|
| 49 |
+
generate_button.click(
|
| 50 |
+
fn=generate_urdu_speech,
|
| 51 |
+
inputs=text_input,
|
| 52 |
+
outputs=audio_output
|
| 53 |
+
)
|
| 54 |
+
|
| 55 |
+
# 4. Launch the Gradio app
|
| 56 |
+
if __name__ == "__main__":
|
| 57 |
+
demo.launch()
|