app.py

import gradio as gr
import wave
import numpy as np
from io import BytesIO
from huggingface_hub import hf_hub_download
from piper import PiperVoice 
from transformers import pipeline

# Import necessary libraries:
# gradio for creating the web interface,
# wave for handling WAV audio format,
# numpy for numerical operations,
# BytesIO for in-memory byte handling,
# huggingface_hub for downloading models from the Hugging Face Hub,
# PiperVoice for the text-to-speech functionality,
# pipeline from transformers for the NSFW classifier.

# Load the NSFW classifier model using Hugging Face's pipeline
nsfw_detector = pipeline("text-classification", model="michellejieli/NSFW_text_classifier")

def synthesize_speech(text):
    # Check for NSFW content using the classifier
    nsfw_result = nsfw_detector(text)
    # Extract the label and score from the result
    label = nsfw_result[0]['label']
    score = nsfw_result[0]['score']

    # First, check if the input text contains NSFW content.
    #nsfw_result = nsfw_detector(text)
    if label == 'NSFW' and score >= 0.95:
        # Download and read the error audio file
        error_audio_path = hf_hub_download(repo_id="DLI-SLQ/speaker_01234", filename="error_audio.wav")
        with open(error_audio_path, 'rb') as error_audio_file:
            error_audio = error_audio_file.read()
        # Return the error audio and a warning message
        return error_audio, "NSFW content detected. Cannot process."

    # If the content is safe, proceed with speech synthesis.
    # Download the model and configuration from Hugging Face Hub.
    model_path = hf_hub_download(repo_id="DLI-SLQ/speaker_01234", filename="speaker__01234_model.onnx")
    config_path = hf_hub_download(repo_id="DLI-SLQ/speaker_01234", filename="speaker__1234_model.onnx.json")
    
    # Load the PiperVoice model for speech synthesis.
    voice = PiperVoice.load(model_path, config_path)
    
    # Create a BytesIO buffer to hold the synthesized WAV file in memory.
    buffer = BytesIO()
    with wave.open(buffer, 'wb') as wav_file:
        # Set WAV file properties: sample rate, bit depth, and mono channel.
        wav_file.setframerate(voice.config.sample_rate)
        wav_file.setsampwidth(2)  # 16-bit
        wav_file.setnchannels(1)  # mono

        # Use the PiperVoice model to synthesize speech from the text.
        voice.synthesize(text, wav_file)

    # Convert the buffer content to a NumPy array, then to bytes for Gradio output.
    buffer.seek(0)
    audio_data = np.frombuffer(buffer.read(), dtype=np.int16)
    return audio_data.tobytes(), None

# Set up the Gradio interface.
with gr.Blocks(theme=gr.themes.Base()) as blocks:
    # Create a user-friendly markdown title and description.
    gr.Markdown("# Text to Speech Synthesizer")
    gr.Markdown("Enter text to synthesize it into speech using models from the State Library of Queensland's collection using Piper.")
    
    # Define Gradio interface components: input textbox, audio output, and output textbox.
    input_text = gr.Textbox(label="Input Text")
    output_audio = gr.Audio(label="Synthesized Speech", type="numpy")
    output_text = gr.Textbox(label="Output Text", visible=True)  # Visible for error messages

    # Define a function to process the input text and produce outputs.
    def process_and_output(text):
        audio, message = synthesize_speech(text)
        if message:
            # If there's a message (e.g., an error message), return None for audio and the message.
            return None, message
        else:
            # Otherwise, return the synthesized audio and None for the message.
            return audio, None

    # Link the processing function to the Gradio interface button.
    submit_button = gr.Button("Synthesize")
    submit_button.click(process_and_output, inputs=input_text, outputs=[output_audio, output_text])

# Launch the Gradio web application.
blocks.launch()