text_to_speech.py

import gc
from functools import partial
import gradio as gr
import torch
from transformers import pipeline
from utils import get_pytorch_device, spaces_gpu, get_torch_dtype


@spaces_gpu
def text_to_speech(model: str, text: str) -> tuple[int, bytes]:
    """Convert text to speech audio using a TTS (Text-to-Speech) model.
    
    This function uses a transformer pipeline to generate speech audio from
    text input. The model is loaded, inference is performed, and then cleaned
    up to free GPU memory.
    
    Args:
        model: Hugging Face model ID to use for text-to-speech.
        text: Input text string to convert to speech.
    
    Returns:
        Tuple containing:
            - int: Sampling rate of the generated audio (e.g., 22050 Hz)
            - bytes: Raw audio data as bytes
    
    Note:
        - Uses safetensors for secure model loading.
        - Automatically selects the best available device (CUDA/XPU/MPS/CPU).
        - Cleans up model and GPU memory after inference.
        - Returns audio in format compatible with Gradio Audio component.
    """
    pytorch_device = get_pytorch_device()
    dtype = get_torch_dtype()
    
    # During inference or evaluation, gradient calculations are unnecessary. Using torch.no_grad()
    # reduces memory consumption by not storing gradients. This can significantly reduce the
    # amount of memory used during the inference phase.
    model_kwargs = {"use_safetensors": True} # Use safetensors to avoid torch.load restriction.
    if dtype is not None:
        model_kwargs["dtype"] = dtype
    
    narrator = pipeline(
        "text-to-speech",
        model,
        device=0 if pytorch_device == "cuda" else -1,
        model_kwargs=model_kwargs
    )
    with torch.no_grad():
        result = narrator(text)
    
    # Clean up GPU memory
    del narrator
    if pytorch_device == "cuda":
        torch.cuda.empty_cache()
    gc.collect()
    return (result["sampling_rate"], result["audio"][0])


def create_text_to_speech_tab(model: str):
    """Create the text-to-speech tab in the Gradio interface.
    
    This function sets up all UI components for text-to-speech generation,
    including input textbox, generate button, and output audio player.
    
    Args:
        model: Hugging Face model ID to use for text-to-speech.
    """
    gr.Markdown("Generate speech from text.")
    text_to_speech_text = gr.Textbox(label="Text")
    text_to_speech_generate_button = gr.Button("Generate")
    text_to_speech_output = gr.Audio(label="Speech")
    text_to_speech_generate_button.click(
        fn=partial(text_to_speech, model),
        inputs=text_to_speech_text,
        outputs=text_to_speech_output
    )