app.py

import numpy as np
import pandas as pd
import os
import librosa as lr
import torch
import torch.nn as nn
import pytorch_lightning as pl
import gradio as gr
from models.model import MFCC_CNN

EMOTIONS = {
    9 : 'Нейтральная, мужской голос',
    7: 'Счастье, мужской голос',
    11: 'Грусть, мужской голос',
    1: 'Злость, мужской голос',
    5: 'Страх, мужской голос',
    3: 'Отвращение, мужской голос',
    13 : 'Удивление, мужской голос',
    8 : 'Нейтральная, женский голос',
    6 : 'Счастье, женский голос',
    10 : 'Грусть, женский голос',
    0 : 'Злость, женский голос',
    4 : 'Страх, женский голос',
    2 : 'Отвращение, женский голос',
    12 : 'Удивление, женский голос'
    }

# LOAD AUDIO
SAMPLE_RATE = 16000
DURATION = 3

# GET MFCC
N_MFCC = 50
WIN_LENGTH = 2048
WINDOW = 'hann'
HOP_LENGTH = 512

PATH = './chekpoint/models-epoch=97-val_loss=2.09.ckpt'
ckpt = torch.load(PATH)

pretrained_model = MFCC_CNN(14)
pretrained_model.load_state_dict(ckpt['state_dict'])
pretrained_model.eval()
pretrained_model.freeze()

def scaler_params():
    with open('./mean_.txt', mode='r') as f:
            mean = f.read()
            mean = mean.split(',')

    mean_final = []
    for number in mean:
        if number != ' ':
            mean_final.append(float(number))

    with open('./scale_.txt', mode='r') as f:
        scale = f.read()
        scale = scale.split(',')

    scale_final = []
    for number in scale:
        if number != ' ':
            scale_final.append(float(number))

    return np.array(mean_final), np.array(scale_final)

def processAudio(audio_file):
    audio, sr = lr.load(audio_file, 
                             duration=DURATION, 
                             sr=SAMPLE_RATE)
        
    signal = np.zeros((int(SAMPLE_RATE*3,)))
    signal[:len(audio)] = audio

    feature_set = []

    mfcc = lr.feature.mfcc(y=signal,
                                    sr=sr,
                                    n_mfcc=N_MFCC,
                                    win_length=WIN_LENGTH,
                                    window=WINDOW,
                                    hop_length=HOP_LENGTH,
                                    )
            
    feature_set = torch.tensor(mfcc, dtype=torch.float)

    feature_set.view(1, 1, 50, 94)
    feature_set = np.reshape(feature_set, newshape=(1,-1))

    mean, scale = scaler_params()
    feature_set = (feature_set - mean)/scale
    
    feature_set = np.reshape(feature_set, newshape=(1, 1, 50, 94))
    feature_set = torch.tensor(feature_set, dtype=torch.float)
    
    prediction = pretrained_model(feature_set,)
    prediction = torch.argmax(prediction)

    return EMOTIONS[prediction.item()]
    
demo = gr.Interface(
    fn=processAudio,
    inputs=gr.Audio(type='filepath'),
    outputs=gr.Label(),
    examples=[
        [os.path.join(os.path.dirname(__file__), "files/03-01-01-01-02-02-01.wav")],
        [os.path.join(os.path.dirname(__file__), "files/03-01-07-01-02-02-01.wav")],
        [os.path.join(os.path.dirname(__file__), "files/03-01-08-02-02-02-01.wav")],
        ],
    )

if __name__ == '__main__':
    demo.launch()