app.py

import gradio as gr
from transformers import pipeline, WhisperProcessor, WhisperForConditionalGeneration
import torch 
import torchaudio
import soundfile as sf

# Load Whisper model and processor
processor = WhisperProcessor.from_pretrained("openai/whisper-large")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large")

# Load the Hugging Face emotion classifier
emotion_classifier = pipeline("text-classification", model="SamLowe/roberta-base-go_emotions", top_k=None)

# Define a function to process audio and analyze emotions
def transcribe_and_analyze(audio_path):
    # Load audio from the provided file
    audio, sample_rate = sf.read(audio_path)
    
    # Resample audio to 16000 Hz if necessary
    print('resample')
    if sample_rate != 16000:
        audio_tensor = torchaudio.functional.resample(torch.tensor(audio), orig_freq=sample_rate, new_freq=16000)
        audio = audio_tensor.numpy()  # Convert back to numpy array
    
    # Process audio with Whisper
    # input_features = model(audio)
    input_features = processor(audio, sampling_rate=16000, return_tensors="pt")
    print(input_features)
    print('trans')
    predicted_ids = model.generate(input_features.input_features)
    
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    print(transcription)

    # Analyze emotions in the transcription
    emotions = emotion_classifier(transcription)
    return transcription, emotions

# Create Gradio interface
interface = gr.Interface(
    fn=transcribe_and_analyze,
    inputs=gr.Audio(type="filepath"),  # Accept audio input
    outputs=[
        gr.Textbox(label="Transcription"),  # Display transcription
        gr.JSON(label="Emotion Analysis")  # Display emotion analysis
    ],
    title="Audio to Emotion Analysis"
)

# Launch the Gradio app
interface.launch()