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

# Load the NSFW classifier model
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)
    label = nsfw_result[0]['label']
    score = nsfw_result[0]['score']

    if label == 'NSFW' and score >= 0.95:
        error_audio_path = hf_hub_download(repo_id="DLI-SLQ/speaker_01234", filename="error_audio.wav")

        # Read the error audio file
        try:
            with wave.open(error_audio_path, 'rb') as error_audio_file:
                frames = error_audio_file.readframes(error_audio_file.getnframes())
                error_audio_data = np.frombuffer(frames, dtype=np.int16).tobytes()
        except Exception as e:
            print(f"Error reading audio file: {e}")
            return None, "Error in processing audio file."

        return error_audio_data, "NSFW content detected. Cannot process."

    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__01234_model.onnx.json")

    voice = PiperVoice.load(model_path, config_path)

    buffer = BytesIO()
    with wave.open(buffer, 'wb') as wav_file:
        wav_file.setframerate(voice.config.sample_rate)
        wav_file.setsampwidth(2)
        wav_file.setnchannels(1)
        voice.synthesize(text, wav_file, sentence_silence=0.75, length_scale=1.2)

    buffer.seek(0)
    audio_data = np.frombuffer(buffer.read(), dtype=np.int16)
    return audio_data.tobytes(), None

# Gradio Interface
with gr.Blocks(theme=gr.themes.Base(),css="footer {visibility: hidden}") as blocks:
    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.  This model uses data from the following collections:  Suzanne Mulligan Oral Histories Archive,  the Peter Gray audio tapes, Five Years On : Toowoomba and Lockyer Valley flash floods: oral history interviews and Our Rocklea: connecting with the heart through story and creativity 2012.")
    input_text = gr.Textbox(label="Input Text")
    submit_button = gr.Button("Synthesize")
    output_audio = gr.Audio(label="Synthesized Speech", type="numpy", show_download_button=False)
    output_text = gr.Textbox(label="Output Text", visible=False)
    


    def process_and_output(text):
        audio, message = synthesize_speech(text)
        if message:
            return audio, message
        else:
            return audio, None

    submit_button.click(process_and_output, inputs=input_text, outputs=[output_audio, output_text])

blocks.launch()