model.py

import onnxruntime as ort
import librosa
import numpy as np
import scipy

SR = 22050
LENGTH_SEC = 1.5

def resample(audio_data, original_sr, target_sr):
    num_samples = int(len(audio_data) * target_sr / original_sr)
    return scipy.signal.resample(audio_data, num_samples)


def load_audio_slices(af):
    audio, sr = librosa.load(af, sr=None)
    if sr != SR:
        audio = resample(audio, sr, SR)

    LENGTH_SAMPLES = int(LENGTH_SEC * SR)

    slices = []
    if len(audio) < LENGTH_SAMPLES:
        padding_needed = LENGTH_SAMPLES - len(audio)
        audio = np.pad(audio, (0, padding_needed), mode='constant')
        slices.append(audio)
    else:
        num_chunks = len(audio) / LENGTH_SAMPLES

        if num_chunks > 2:
            num_chunks = 5
        elif num_chunks > 1.5:
            num_chunks = 3
        elif num_chunks > 1:
            num_chunks = 2

        end = len(audio) - LENGTH_SAMPLES
        idxs_split = np.arange(0, end, end // num_chunks, dtype=int)
        for idx_split in idxs_split:
            sl = slice(idx_split, idx_split+LENGTH_SAMPLES)
            slices.append(audio[sl])

    slices = np.vstack(slices)

    return slices / np.max(slices, axis=1)[:, np.newaxis]

class CnnVoiceClassifier:
    def __init__(self):
        self.session = ort.InferenceSession('model.onnx')
        self.input_name = self.session.get_inputs()[0].name
        self.output_name = self.session.get_outputs()[0].name

    def inference(self, audio_path):
        audio = load_audio_slices(audio_path)
        input_feed = {self.input_name: np.expand_dims(audio, axis=-1)}
        outputs = self.session.run([self.output_name], input_feed)
        probs = outputs[0].flatten()
        w = np.abs((probs - 0.5)*2)
        final_prob = np.average(probs, weights=w)

        return {'Male': final_prob, 'Female': 1-final_prob}