This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on Librispeech-clean-100 for gender recognition. It achieves the following results on the evaluation set:

Loss: 0.0061
F1: 0.9993

Compute your inferences

python

import os
import random
from glob import glob
from typing import List, Optional, Union, Dict

import tqdm
import torch
import torchaudio
import numpy as np
import pandas as pd
from torch import nn
from torch.utils.data import DataLoader
from torch.nn import functional as F
from transformers import (
    AutoFeatureExtractor,
    AutoModelForAudioClassification,
    Wav2Vec2Processor
)

class CustomDataset(torch.utils.data.Dataset):
    def __init__(
        self,
        dataset: List,
        basedir: Optional[str] = None,
        sampling_rate: int = 16000,
        max_audio_len: int = 5,
    ):
        self.dataset = dataset
        self.basedir = basedir

        self.sampling_rate = sampling_rate
        self.max_audio_len = max_audio_len

    def __len__(self):
        """
        Return the length of the dataset
        """
        return len(self.dataset)

    def __getitem__(self, index):
        if self.basedir is None:
            filepath = self.dataset[index]
        else:
            filepath = os.path.join(self.basedir, self.dataset[index])

        speech_array, sr = torchaudio.load(filepath)

        if speech_array.shape[0] > 1:
            speech_array = torch.mean(speech_array, dim=0, keepdim=True)

        if sr != self.sampling_rate:
            transform = torchaudio.transforms.Resample(sr, self.sampling_rate)
            speech_array = transform(speech_array)
            sr = self.sampling_rate

        len_audio = speech_array.shape[1]

        # Pad or truncate the audio to match the desired length
        if len_audio < self.max_audio_len * self.sampling_rate:
            # Pad the audio if it's shorter than the desired length
            padding = torch.zeros(1, self.max_audio_len * self.sampling_rate - len_audio)
            speech_array = torch.cat([speech_array, padding], dim=1)
        else:
            # Truncate the audio if it's longer than the desired length
            speech_array = speech_array[:, :self.max_audio_len * self.sampling_rate]

        speech_array = speech_array.squeeze().numpy()

        return {"input_values": speech_array, "attention_mask": None}


class CollateFunc:
    def __init__(
        self,
        processor: Wav2Vec2Processor,
        padding: Union[bool, str] = True,
        pad_to_multiple_of: Optional[int] = None,
        return_attention_mask: bool = True,
        sampling_rate: int = 16000,
        max_length: Optional[int] = None,
    ):
        self.sampling_rate = sampling_rate
        self.processor = processor
        self.padding = padding
        self.pad_to_multiple_of = pad_to_multiple_of
        self.return_attention_mask = return_attention_mask
        self.max_length = max_length

    def __call__(self, batch: List[Dict[str, np.ndarray]]):
        # Extract input_values from the batch
        input_values = [item["input_values"] for item in batch]

        batch = self.processor(
            input_values,
            sampling_rate=self.sampling_rate,
            return_tensors="pt",
            padding=self.padding,
            max_length=self.max_length,
            pad_to_multiple_of=self.pad_to_multiple_of,
            return_attention_mask=self.return_attention_mask
        )

        return {
            "input_values": batch.input_values,
            "attention_mask": batch.attention_mask if self.return_attention_mask else None
        }


def predict(test_dataloader, model, device: torch.device):
    """
    Predict the class of the audio
    """
    model.to(device)
    model.eval()
    preds = []

    with torch.no_grad():
        for batch in tqdm.tqdm(test_dataloader):
            input_values, attention_mask = batch['input_values'].to(device), batch['attention_mask'].to(device)

            logits = model(input_values, attention_mask=attention_mask).logits
            scores = F.softmax(logits, dim=-1)

            pred = torch.argmax(scores, dim=1).cpu().detach().numpy()

            preds.extend(pred)

    return preds


def get_gender(model_name_or_path: str, audio_paths: List[str], label2id: Dict, id2label: Dict, device: torch.device):
    num_labels = 2

    feature_extractor = AutoFeatureExtractor.from_pretrained(model_name_or_path)
    model = AutoModelForAudioClassification.from_pretrained(
        pretrained_model_name_or_path=model_name_or_path,
        num_labels=num_labels,
        label2id=label2id,
        id2label=id2label,
    )

    test_dataset = CustomDataset(audio_paths, max_audio_len=5)  # for 5-second audio

    data_collator = CollateFunc(
        processor=feature_extractor,
        padding=True,
        sampling_rate=16000,
    )

    test_dataloader = DataLoader(
        dataset=test_dataset,
        batch_size=16,
        collate_fn=data_collator,
        shuffle=False,
        num_workers=2
    )

    preds = predict(test_dataloader=test_dataloader, model=model, device=device)

    return preds

model_name_or_path = "alefiury/wav2vec2-large-xlsr-53-gender-recognition-librispeech"

audio_paths = [] # Must be a list with absolute paths of the audios that will be used in inference
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

label2id = {
    "female": 0,
    "male": 1
}

id2label = {
    0: "female",
    1: "male"
}

num_labels = 2

preds = get_gender(model_name_or_path, audio_paths, label2id, id2label, device)

Training and evaluation data

The Librispeech-clean-100 dataset was used to train the model, with 70% of the data used for training, 10% for validation, and 20% for testing.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 3e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 16
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_epochs: 1
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	F1
0.002	1.0	1248	0.0061	0.9993

Framework versions

Transformers 4.28.0
Pytorch 2.0.0+cu118
Tokenizers 0.13.3