Speaker Identification with E-TDNN embeddings on Voxceleb

This repository provides a pretrained E-TDNN model (x-vector) using SpeechBrain. The system can be used to extract speaker embeddings as well. Since we can't find any resource that has SpeechBrain or HuggingFace compatible checkpoints that has only been trained on VoxCeleb2 development data, so we decide to pre-train an E-TDNN system from scratch.

Pipeline description

This system is composed of an E-TDNN model (x-vector). It is a combination of convolutional and residual blocks. The embeddings are extracted using temporal statistical pooling. The system is trained with Additive Margin Softmax Loss.

We use FBank (16kHz, 25ms frame length, 10ms hop length, 80 filter-bank channels) as the input features. It was trained using initial learning rate of 0.001 and batch size of 512 with linear scheduler for 40 epochs on 4 A100 GPUs. We employ additive noises and reverberation from MUSAN and RIR datasets to enrich the supervised information. The pre-training progress takes approximately seven days for the E-TDNN model.

Performance

VoxCeleb1-O is the original verification test set from VoxCeleb1 consisting of 40 speakers. All speakers with names starting with "E" are reserved for testing. VoxCeleb1-E uses the entire VoxCeleb1 dataset, covering 1251 speakers. VoxCeleb1-H is a hard version of evaluation set consisting of 552536 pairs with 1190 speakers with the same nationality and gender. There are 18 nationality-gender combinations each with at least 5 individuals.

Splits	Backend	S-norm	EER(%)	minDCF(0.01)
VoxCeleb1-O	cosine	no	1.91	0.20
VoxCeleb1-E	cosine	no	TBD	TBD
VoxCeleb1-H	cosine	no	TBD	TBD

VoxCeleb1-O: includes 37611 test pairs with 40 speakers.
VoxCeleb1-E: includes 579818 test pairs with 1251 speakers.
VoxCeleb1-H: includes 550894 test pairs with 1190 speakers.

Compute the speaker embeddings

The system is trained with recordings sampled at 16kHz (single channel).

import torch
import torchaudio
from speechbrain.pretrained.interfaces import Pretrained
from speechbrain.pretrained import EncoderClassifier


class Encoder(Pretrained):

    MODULES_NEEDED = [
        "compute_features",
        "mean_var_norm",
        "embedding_model"
    ]

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def encode_batch(self, wavs, wav_lens=None, normalize=False):
        # Manage single waveforms in input
        if len(wavs.shape) == 1:
            wavs = wavs.unsqueeze(0)

        # Assign full length if wav_lens is not assigned
        if wav_lens is None:
            wav_lens = torch.ones(wavs.shape[0], device=self.device)

        # Storing waveform in the specified device
        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
        wavs = wavs.float()

        # Computing features and embeddings
        feats = self.mods.compute_features(wavs)
        feats = self.mods.mean_var_norm(feats, wav_lens)
        embeddings = self.mods.embedding_model(feats, wav_lens)
        if normalize:
            embeddings = self.hparams.mean_var_norm_emb(
                embeddings,
                torch.ones(embeddings.shape[0], device=self.device)
            )
        return embeddings


classifier = Encoder.from_hparams(
    source="yangwang825/etdnn-vox2"
)
signal, fs = torchaudio.load('spk1_snt1.wav')
embeddings = classifier.encode_batch(signal)
>>> torch.Size([1, 1, 192])

We will release our training results (models, logs, etc) shortly.

References

Ravanelli et al., SpeechBrain: A General-Purpose Speech Toolkit, 2021
Snyder et al., The JHU Speaker Recognition System for the VOiCES 2019 Challenge, 2019