Analysis of i-vector Length Normalization in Speaker Recognition Systems

By Daniel Garcia-Romero, Carol Y. Espy-Wilson. 2011

1,065 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Signal Processing

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

End-to-end text-dependent speaker verification (2015, background): “The second baseline is based on [ 25] with 150 eigenvoices.”
A Novel Learnable Dictionary Encoding Layer for End-to-End Language Identification (2018, methodology): “As is typical in conventional GMM Supervector/ivector, the resulting vectors are normalized using the length normalization [24].”
A Divide-and-Conquer Approach for Language Identification Based on Recurrent Neural Networks (2016, methodology): “Here, the PLDA (Probabilistic Linear Discriminant Analysis) technique [27], which is also commonly used for speaker verification [24, 25] or gender identification [28], was used.”
Voices Obscured in Complex Environmental Settings (VOICES) corpus (2018, methodology): “The model used is a Universal Background Model (UBM) identity vector (i-vector) based system [13, 14], with a probabilistic linear discriminant analysis (PLDA) [15] as back-end classiﬁer.”
SpeechBrain: A General-Purpose Speech Toolkit (2021, background): “Furthermore, SpeechBrain provides traditional Probabilistic Linear Discriminant Analysis (PLDA) for speaker discrimination [93, 94].”
AutoSpeech: Neural Architecture Search for Speaker Recognition (2020, positive): “, MFCCs) as the speaker embedding [24, 25], but these systems have been superseded by systems based on “i-vectors” [26, 27, 28].”
Speaker recognition based on deep learning: An overview (2020, methodology): “The baselin is a traditional i-vector syst m that is based on the GMM-UBM Kaldi recipe described in [11].”
Front-End Factor Analysis For Speaker Verification (2018, methodology): “However, by applying to the i-vectors a non-linear transformation called “length normalization” [16], the performance of Gaussian PLDA becomes similar with the one of heavy-tailed PLDA (without applying length normalization).”
Attention-Based Models for Text-Dependent Speaker Verification (2017, methodology): “I-vector [3] based systems in combination with verification back-ends such as Probabilistic Linear Discriminant Analysis (PLDA) [4] have been the dominating paradigm of SV in previous years.”
Self-Attentive Speaker Embeddings for Text-Independent Speaker Verification (2018, background): “Over the years, the combination of i-vectors [1] and probabilistic linear discriminant analysis (PLDA) [2] has been the dominant approach for text-independent SV tasks [3, 4, 5].”
Probabilistic linear discriminant analysis of i-vector posterior distributions (2013, methodology): “In this work we consider only PLDA with Gaussian priors, because these models have shown to be accurate and effective with respect to more computational expensive models such as the Heavy– Tailed PLDA [4] [7].”
The IBM 2016 Speaker Recognition System (2016, methodology): “The Eigenvoice subspace in the PLDA model is assumed full-rank.”
A Review of Speaker Diarization: Recent Advances with Deep Learning (2021, methodology): “As discussed in [167], the RRNN-based distance estimation exhibits consistent improvements in its performance when compared with the more traditional distance estimation approaches such as the cosine distance [30] or the PLDA-based [133] distance.”
Phonetically-constrained PLDA modeling for text-dependent speaker verification with multiple short utterances (2013, methodology): “Thus, several normalization techniques have been proposed to make the i-vector distribution closer to the Gaussian assumption [9, 10].”
Text-dependent speaker recognition using PLDA with uncertainty propagation (2013, methodology): “Length normalization is applied, i.e. i ← √ d i k i k , where the multiplicative term √ d is placed in order to preserve the scaling of the original i -vector space, [12].”
Spoofing and countermeasures for speaker verification: A survey (2015, methodology): “The normalisation of i-vectors to lie on a unit sphere is also popular as a pre-processing technique for the PLDA back-end (Garcia-Romero and Espy-Wilson, 2011).”
The NIST 2014 Speaker Recognition i-vector Machine Learning Challenge (2014, methodology): “There was a large performance gap observed between the baseline and the oracle PLDA system for both the progress and evaluation sets.”