app.py

import gradio as gr
import numpy as np
import torch
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan

# Load the model and vocoder
checkpoint = "microsoft/speecht5_tts"
processor = SpeechT5Processor.from_pretrained(checkpoint)
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

# Speaker embeddings for male and female
speaker_embeddings = {
    "male": "speaker/cmu_us_bdl_arctic-wav-arctic_a0009.npy",
    "female": "speaker/cmu_us_slt_arctic-wav-arctic_a0508.npy"
}

# Function to generate speech
def text_to_speech(text, gender):
    if len(text.strip()) == 0:
        return (16000, np.zeros(0).astype(np.int16))

    inputs = processor(text=text, return_tensors="pt")

    # Truncate input if too long
    input_ids = inputs["input_ids"]
    input_ids = input_ids[..., :model.config.max_text_positions]

    # Load speaker embedding based on gender selection
    speaker_embedding_path = speaker_embeddings[gender]
    speaker_embedding = np.load(speaker_embedding_path)
    speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)

    # Generate speech
    speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
    speech = (speech.numpy() * 32767).astype(np.int16)

    return (16000, speech)

# Create the Gradio interface
iface = gr.Interface(
    fn=text_to_speech,
    inputs=[
        gr.Textbox(label="Enter Text"),
        gr.Radio(["male", "female"], label="Select Voice Gender")  # Gender selection
    ],
    outputs=gr.Audio(label="Generated Speech"),
    title="Text-to-Speech Bot",
    description="Enter text and select a voice gender to generate speech."
)

# Launch the interface
iface.launch()