Attentive Statistics Pooling for Deep Speaker Embedding

By Koji Okabe, Takafumi Koshinaka, K. Shinoda. Interspeech, 2018

633 citations recorded by the source corpus.

This paper proposes attentive statistics pooling for deep speaker embedding in text-independent speaker verification. In conventional speaker embedding, frame-level features are averaged over all the frames of a single utterance to form an utterance-level feature. Our method utilizes an attention mechanism to give different weights to different frames and generates not only weighted means but also weighted standard deviations. In this way, it can capture long-term variations in speaker characteristics more effectively. An evaluation on the NIST SRE 2012 and the VoxCeleb data sets shows that it reduces equal error rates (EERs) from the conventional method by 7.5% and 8.1%, respectively.

Subfields: Artificial Intelligence, Signal Processing

View Paper: https://www.semanticscholar.org/paper/4a64008cd21488071ac418aad32ce7129408edb4

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Signal Processing

Citations in the corpus

Deep Speaker Embedding Extraction with Channel-Wise Feature Responses and Additive Supervision Softmax Loss Function (2019, methodology): “[5] TDNN (x-vector) Softmax VoxCeleb1 Kaldi 3.”
The Idlab Voxsrc-20 Submission: Large Margin Fine-Tuning and Quality-Aware Score Calibration in DNN Based Speaker Verification (2020, background): “The statistics pooling layer that maps the variable length input to a fixedlength representation can be enhanced through a temporal attention mechanism [6, 7].”
Speaker recognition based on deep learning: An overview (2020, methodology): “ss of the orthogonality regularization by TDNN. Generally, TDNN has been frequently used as a framework to study other key components of the deep embedding models, such as the temporal pooling layers [84, 85] and objective functions [86, 87, 88]. The TDNN structure has also been intensively improved”
Exploring wav2vec 2.0 on speaker verification and language identification (2020, neutral): “41 attentive statistics [28] 3.”
Unsupervised Adaptation of Acoustic Models for ASR Using Utterance-Level Embeddings from Squeeze and Excitation Networks (2019, methodology): “Recently, attention-based pooling is used to generate speaker embeddings for speaker recognition and verification tasks [35], [38].”
The SpeakIn System for VoxCeleb Speaker Recognition Challange 2021 (2021, background): “An attention mechanism [16] is introduced to calculate weighted statistics of the outputs of the backbone.”
ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification (2020, methodology): “The attention system [7] allows the model to select the frames it deems relevant.”
Clova Baseline System for the VoxCeleb Speaker Recognition Challenge 2020 (2020, methodology): “We refer to this configuration as H / ASP in the results.”
NANSY++: Unified Voice Synthesis with Neural Analysis and Synthesis (2022, methodology): “…timbre tokens - using timbre encoder f θ T . f θ T takes mel spectrogram as an input and produces global timbre embedding g , a single vector summarized with attentive statistical time-pooling (Okabe et al., 2018), which is expected to capture overall timbral characteristics within an utterance.”
Voxceleb: Large-scale speaker verification in the wild (2020, methodology): “Typically, such systems are trained end-to-end for classi ﬁ cation with a softmax loss (Okabe et al., 2018) or one of its variants, such as the angular softmax (Cai et al., 2018c).”
RawNet: Advanced end-to-end deep neural network using raw waveforms for text-independent speaker verification (2019, methodology): “[25] MFCC x-vector PLDA Softmax 1500 Heavy 3.”
Utterance-level Aggregation for Speaker Recognition in the Wild (2019, methodology): “[15, 13, 19] do not report results on the VoxCeleb-E and VoxCeleb-H) [3] test sets.”
In defence of metric learning for speaker recognition (2020, methodology): “A range of encoders have been proposed to aggregate frame-level informations into utterance-level embeddings, from simple averaging [1] to statistical pooling [4, 7] and dictionary-based encodings [17, 34].”
Improved RawNet with Filter-wise Rescaling for Text-independent Speaker Verification using Raw Waveforms (2020, methodology): “However, applying an attention mechanism to speaker verification is more focused on attentive pooling than applying such mechanisms to intermediate feature maps in image domain tasks [7, 16, 17].”
End-to-End anti-spoofing with RawNet2 (2020, methodology): “Experiments using the VoxCeleb1 database show that RawNet1 gives an 11% relative improvement in terms of an equal error rate (EER) metric compared to an i-vector based system [16], and is also shown to be competitive with those of an x-vector system [17].”