app.py

from transformers import pipeline
# Load model directly
from transformers import AutoProcessor, AutoModelForTextToWaveform

import gradio as gr


processor = AutoProcessor.from_pretrained("suno/bark-small")
# Load audio classification model
audio_classifier = pipeline(
    "audio-classification", model="HareemFatima/distilhubert-finetuned-stutterdetection"
)

# Load text-to-speech model
tts_processor = AutoProcessor.from_pretrained("suno/bark-small")
tts_model = AutoModelForTextToWaveform.from_pretrained("suno/bark-small")

# Define therapy text for different stutter types (replace with your specific therapy content)
therapy_text = {
    "Normal Speech": "Your speech sounds great! Keep practicing!",
    "Blocking": "Take a deep breath and try speaking slowly. You can do it!",
    "Prolongation": "Focus on relaxing your mouth muscles and speaking smoothly.",
    # Add more stutter types and therapy text here
}


def predict_and_synthesize(audio):
    """Predicts stutter type and synthesizes speech with therapy text.

    Args:
        audio (bytes): Audio data from the user.

    Returns:
        tuple: A tuple containing the predicted stutter type (string) and synthesized speech (bytes).
    """

    # Classify stuttering type using audio classification model
    prediction = audio_classifier(audio)
    stutter_type = prediction[0]["label"]

    # Retrieve therapy text based on predicted stutter type
    therapy = therapy_text.get(stutter_type, "General therapy tip: Practice slow, relaxed speech.")

    # Generate synthesized speech with the therapy text
    synthesized_speech = tts_model.generate(
        tts_processor(therapy, return_tensors="pt").input_ids
    )[0].squeeze().cpu().numpy()

    return stutter_type, synthesized_speech


# Create Gradio interface
interface = gr.Interface(
    fn=predict_and_synthesize,
    inputs="microphone",
    outputs=["text", "audio"],
    title="Stuttering Therapy Assistant",
    description="This app helps you identify stuttering types and provides personalized therapy suggestions. Upload an audio clip, and it will analyze the speech and generate audio with relevant therapy tips.",
)

interface.launch(debug=False)