app.py

import gradio as gr
import librosa
import numpy as np
import torch

from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan

checkpoint = "burraco135/speecht5_finetuned_voxpopuli_it"
processor = SpeechT5Processor.from_pretrained(checkpoint)
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

def predict(text, speaker):

    speaker_embedding = np.load("speaker_0_embeddings.npy")
    
    inputs = processor(text=text, return_tensors="pt")

    # limit input length
    input_ids = inputs["input_ids"]
    input_ids = input_ids[..., :model.config.max_text_positions]

    speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)

    speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)

    speech = (speech.numpy() * 32767).astype(np.int16)
    return (16000, speech)

gr.Interface(
    fn=predict,
    inputs=[
        gr.Text(label="Input Text"),
    ],
    outputs=[
        gr.Audio(label="Generated Speech", type="numpy"),
    ]
).launch()