This paper describes the experiments conducted for spoken web search (SWS) at MediaEval 2013 evaluations. A conventional approach is to train a multi-layer perceptron using high resource languages and then use it in the low resource scenario. However, phone posteriorgrams have been found to under-perform when the language they were trained on di ers from the target language. In this paper, we use bottle-neck features derived from MLP to generate Gaussian posteriorgrams. We also use a variant of dynamic time warping (DTW) based technique which exploits the redundancy in speech signal and thus averages the successive Gaussian posteriorgrams to reduce the length of the spoken query and spoken reference.
Gaussian and phone posteriorgrams are a popular feature
representation for query-by-example spoken term detection
(QbE-STD). Gaussian posteriorgrams are typically trained
in an unsupervised manner often referred to as zero-resource
scenario, whereas, phone posteriorgrams are obtained by
training a multi-layer perceptron (MLP) in a supervised
manner. For low/zero resource languages, an MLP is trained
on high resource languages and then it is used in the low
resource scenario. However, phone posteriorgrams have been
found to under-perform when the language they were trained
on di ers from the target language. These MLP classi er
outputs, though capture acoustic phonetic properties of a
speech signal, are not su cient as a feature representation.
This is because the language used for training MLP is not
enough to capture the complete acoustic characteristics of
the multi-lingual data. To utilize this complimentary
information captured, we derive features from an MLP for
obtaining Gaussian posteriorgrams. A similar kind of feature
representation has been explored in paper [
An alternative representation for phone posteriorgrams are the articulatory features (AFs). AFs are a better representation as they are more language universal than phones.
This paper describes the experiments conducted for
spoken web search (SWS) at MediaEval 2013 [
We use a three step process to generate the features for
QbE-STD: (a) Extracting speech parameters such as
frequency domain linear prediction (FDLP) [
In [
A 25 ms window length with 10 ms shift was considered to extract 13 dimensional features along with delta and acceleration coe cients for FDLP. An all-pole model of order 160 poles/sec and 37 lter banks are considered to extract FDLP. 2.1
In this paper, we train phone and AF MLPs using labelled
Telugu database ( 24 hours) consisting of 49 phones [
voicing short, long, diphthong
high, mid, low front, central, back
rounding stop, fricative, a ricative
nasal, approximant velar, alveolar, palatal, labial, dental aspiration silence
The articulatory features (AFs) used in this work
represent the characteristics of speech production process, which
include vowel properties, place of articulation, manner of
articulation, etc. We modi ed the AFs described in [
Gaussian posteriorgrams are computed by training a
Gaussian mixture model (GMM) on the spoken data and the
posterior probability obtained from each Gaussian is used to
represent the speech parameters. The number of Gaussians
represent the approximate number of acoustic units present
in the spoken data. We computed Gaussian posteriorgrams
as described in [
We used a variant of DTW-based approach, referred to as
non-segmental DTW (NS-DTW), for obtaining the search
results [
To improve the computational performance, we reduce the query and reference Gaussian posteriorgrams vectors before performing search. Given a reduction factor 2 N, a window of size is considered over the posteriorgram features and a mean is computed. The window is then shifted by and another mean vector is computed. The posteriorgram vectors are replaced with the reduced number of posteriorgram features during this process. The averaging of Gaussian posteriorgrams also reduce the amount of memory required to compute the similarity matrix. In a conventional approach the space complexity required to compute the similarity matrix between a query and reference is of order O(mnd2) where m,n are the length of reference and query and d is the dimension of the feature vector. The averaging of Gaussian posteriorgrams will reduce the space complexity to an order of O( mn2d2 ) . 4.
In this work we have used the bottle-neck features obtained from phone and articulatory MLPs. We have shown that these BN features perform better than the conventional Gaussian posteriorgrams computed from FDLP. This motivates us to build models using high resource languages and use it in the low resource scenario. 5.