Scribd Logo
Upload

Welcome to Scribd!

  • 177 views

    Usc Poster

    Uploaded by

    api-332129590
    This document proposes a novel i-vector based convolutive non-negative matrix factorization (CNMF) approach for noise robust automatic speech recognition. It aims to learn robust speech representations that are invariant to speaker environment and recording conditions. The approach uses CNMF to represent spectrograms as a product of a basis speech dictionary and time activations. It then learns the total variability matrix to model noise and channel variations in the speech dictionary, leading to features invariant to noisy environments. The method is evaluated on the Aurora 4 noisy speech database, and results show the i-vector CNMF approach outperforms traditional filterbank features for noise robust ASR. Future work includes hyperparameter optimization and integrating the features into a GMM-DNN ASR

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Usc Poster

    Uploaded by

    api-332129590
    177 views1 page
    This document proposes a novel i-vector based convolutive non-negative matrix factorization (CNMF) approach for noise robust automatic speech recognition. It aims to learn robust speech representations that are invariant to speaker environment and recording conditions. The approach uses CNMF to represent spectrograms as a product of a basis speech dictionary and time activations. It then learns the total variability matrix to model noise and channel variations in the speech dictionary, leading to features invariant to noisy environments. The method is evaluated on the Aurora 4 noisy speech database, and results show the i-vector CNMF approach outperforms traditional filterbank features for noise robust ASR. Future work includes hyperparameter optimization and integrating the features into a GMM-DNN ASR

    Original Title

    usc poster

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document proposes a novel i-vector based convolutive non-negative matrix factorization (CNMF) approach for noise robust automatic speech recognition. It aims to learn robust speech representations that are invariant to speaker environment and recording conditions. The approach uses CNMF to represent spectrograms as a product of a basis speech dictionary and time activations. It then learns the total variability matrix to model noise and channel variations in the speech dictionary, leading to features invariant to noisy environments. The method is evaluated on the Aurora 4 noisy speech database, and results show the i-vector CNMF approach outperforms traditional filterbank features for noise robust ASR. Future work includes hyperparameter optimization and integrating the features into a GMM-DNN ASR

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    177 views1 page

    Usc Poster

    Uploaded by

    api-332129590
    This document proposes a novel i-vector based convolutive non-negative matrix factorization (CNMF) approach for noise robust automatic speech recognition. It aims to learn robust speech representations that are invariant to speaker environment and recording conditions. The approach uses CNMF to represent spectrograms as a product of a basis speech dictionary and time activations. It then learns the total variability matrix to model noise and channel variations in the speech dictionary, leading to features invariant to noisy environments. The method is evaluated on the Aurora 4 noisy speech database, and results show the i-vector CNMF approach outperforms traditional filterbank features for noise robust ASR. Future work includes hyperparameter optimization and integrating the features into a GMM-DNN ASR

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    You are on page 1of 1

    A NOVEL I-VECTOR BASED CNMF APPROACH TOWARDS NOISE ROBUST ASR

    Kunal Dhawan1 , Colin Vaz2 , Ruchir Travadi2 and Shrikanth Narayanan2


    1
    Department of Electronics and Electrical Engineering, Indian InsPtute of Technology GuwahaP, GuwahaP, Assam, India
    2 Signal Analysis and InterpretaPon Lab, University of Southern California, Los Angeles, CA 90089

    Motivation Proposed Feature GeneraPon Algorithm Dataset DescripPon


    ➢ Problem: Commonly used acoustic features for Automatic Step 1: Learning UBM Speech Dictionary Used the Aurora 4 database which has the following characteristics:
    7 noise types: airport, babble, car, clean, restaurant, street, train
    j jj Speech Recognition are not robust to noise 2 microphone position conditions: near field, far field
    Degrades ASR performance in noisy conditions Training set of 7138 utterances, each approximately 8 seconds
    ➢ Aim: Learn robust representation of speech which is invariant clclong ( noise is not labeled)
    k to speaker environment and recording conditions Test set of 330 utterances for each noise+channel type (∴
    7*2=14 cc different conditions)
    Background Result and Discussion
    1) Non-Negative Matrix Factorization (NMF) Step 2: Calculating the 0th and 1st order sufficient statistics ,
    estimating the mean and covariance supervectors
    The i-vector approach presented is very effective in modeling the p
    An algorithm to approximate a non-negative matrix V
    ppnoise type:
    (∈ℝ≥0,M×N) as a product of two non-negative matrices W
    (∈ℝ≥0,M×R) and H (∈ℝ≥0,R×N), where R≤M.
    To minimize the error of reconstruction, decomposition is
    done according to an adapted Kullback-Leibler divergence
    cost metric:
    123
    #(%| '×) = ∑,-(%,- log 4×5 23
    − %,- + '×) ,- )
    Step 3: Learning the Total Variability Matrix
    The T matrix captures the variations in the training set about the UBM and
    hence defines the space where the speech dictionaries will adapt
    This idea can be used in ASR to represent a spectrogram as a
    product of a basis dictionary and corresponding time Step 4: Calculating the adapted Dictionaries for the training
    activations: set and hence find corresponding time activations Learning the noise and channel variations in the speech dictionary
    ppitself leads to features which are invariant of the noisy
    lllllenvironment:

    Spectrogram (V) Speech Dict. (W) Time Ac9va9ons (H)


    Spectro-temporal repn. of Learns a set of basis Chooses the appropriate
    a given speech sample elements having different basis elements for each Pme
    frequency characteristics frame in a given u]erance

    2) Total Variability approach


    Model the variations in a dataset as low-dimensional vector
    Car Noise Airport Noise Clean Speech Car Noise

    Complete Data distribution Utterance Specific Data


    (UBM) Distribution (means slightly shifted
    from UBM) Babble Noise

    Feature Extraction Step:

    Babble Noise Airport Noise

    9 = : + ;×< i-Vector (a random


    vector having a standard Feature corresponding to Conclusion and Future Work
    Speaker & Channel dependent given utterance
    normal distribution)
    dicPonary supervector ⭐ Current system outperforms the classical filterbank features.
    Step 5: Extract features for the test set using the same method ⭐ Currently performing a grid search over the hyperparameters
    UBM supervector (Speaker & Channel Total Variability Matrix (rectangular
    independent dicPonary supervector) matrix of low rank which models the as in step 4 llll(sparsity of the dicPonary , number of basis elements to choose)
    variation space) lllllwith the final aim of building GMM-DNN ASR system.

    You might also like

    pFad - Phonifier reborn

    Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

    Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


    Alternative Proxies:

    Alternative Proxy

    pFad Proxy

    pFad v3 Proxy

    pFad v4 Proxy