app.py

import opensmile
import joblib
import wave
import datetime
import os
import pandas as pd
from sklearn.preprocessing import StandardScaler


from base64 import b64decode

import gradio as gr


model_path = "RF_emobase_20_model_top1_score0.6863_20231207_1537.joblib"
model = joblib.load(model_path)

def extract_features(audio_path):
    smile = opensmile.Smile(
    #feature_set=opensmile.FeatureSet.GeMAPSv01b,
    feature_set=opensmile.FeatureSet.emobase,
    feature_level=opensmile.FeatureLevel.Functionals,
    )
    feature_df = smile.process_files(audio_path)
    output_features = ['F0env_sma_de_amean', 'lspFreq_sma_de[5]_linregc1', 'mfcc_sma[3]_linregc1', 'lspFreq_sma[6]_quartile1', 'lspFreq_sma_de[6]_linregerrQ', 'lspFreq_sma_de[6]_maxPos', 'lspFreq_sma_de[6]_iqr2-3', 'lspFreq_sma_de[7]_minPos', 'lspFreq_sma_de[4]_linregc1', 'lspFreq_sma_de[6]_linregerrA', 'lspFreq_sma_de[6]_linregc2', 'lspFreq_sma[5]_amean', 'lspFreq_sma_de[6]_iqr1-2', 'mfcc_sma[1]_minPos', 'mfcc_sma[4]_linregc1', 'mfcc_sma[9]_iqr2-3', 'lspFreq_sma[5]_kurtosis', 'lspFreq_sma_de[3]_skewness', 'mfcc_sma[3]_minPos', 'mfcc_sma[12]_linregc1']
    df = pd.DataFrame(feature_df.values[0], index=feature_df.columns)
    df = df[df.index.isin(output_features)]
    df = df.T
    scaler = StandardScaler()
    feature = scaler.fit_transform(df)
    print(df.shape)

    return feature

def main(input):
  # openSMILEで特徴量抽出
  feature_vector = extract_features([input])

  # ロードしたモデルで推論
  prediction = model.predict(feature_vector)
  #print(f"Prediction: {prediction}")
  return prediction

gr.Interface(
    title = 'Question Classifier Model',
    fn = main,
    inputs=[
        gr.Audio(sources=["microphone","upload"], type="filepath")
    ],
    outputs=[
        "textbox"
    ],
    live=True
    ).launch(debug=True)