app.py

import gradio as gr
from transformers import pipeline
import json
import os

data_dict = {}
with open('./results_classification/file.json', 'r') as file:
    data = json.load(file)    
intents_dict = data
tokenizer = AutoTokenizer.from_pretrained("roberta-base")
model = AutoModelForSequenceClassification.from_pretrained("./results_classification/checkpoint-1890/")

def preprocess(text):
    inputs = tokenizer(text, padding=True, truncation=True, return_tensors="pt")
    return inputs

def postprocess(outputs):
    logits = outputs.logits
    predicted_labels = logits.argmax(dim=1).tolist()
    return predicted_labels

def predict(text):
    inputs = preprocess(text)
    with torch.no_grad():
        outputs = model(**inputs)
    predicted_labels = postprocess(outputs)
    ans = intents_dict[predicted_labels[0]]
    return ans

from transformers import pipeline

p = pipeline(model="openai/whisper-medium")

def transcribe(audio):
    t = p(audio)['text']
    ans = predict(t)
    return ans


get_intent = gr.Interface(fn = transcribe,
                          inputs=gr.Audio(source="microphone", type="filepath"),
                          outputs="text").launch()