Citation Context

... the length of the segment. Another approach to discriminative keyword spotting uses kernel machines and large margin classifiers with a set of feature functions consistent with hybrid ANN/HMM models =-=[8]-=-. The approach presented in this paper attempts to model full keywords in the sequence data stream, while previous discriminative keyword spotting systems are based on sub-word units (phones). It is b...

Citation Context

...(EM) algorithm to local maxima, the assumption of conditional independence of the observations, and the fact that HMMs do not directly maximize the keyword detection rate. For these reasons we follow =-=[3]-=- in using a supervised, discriminative approach to keyword spotting, that does not require the use of HMMs. In general, discriminative learing algorithms are likely to outperform generative models suc...

Citation Context

... recordings. The phrases to be detected, also refered to as queries, typically consist of 4– 8 words and have a duration of about 1–3 seconds. Classical approaches to key-word and key-phrase spotting =-=[1]-=- as well as to utterance verification [2] employ a supervised approach where models such as HMMs or SVMs are trained in a preprocessing step. However, in several applications such kind of training is ...

Citation Context

...g the models. Apart from the numerous HMM based approaches, more unconventional keyword spotting strategies, such as applying recurrent neural networks [5] or using discriminative learning procedures =-=[6]-=-, [7] have been developed. The latter technique non-linearly maps speech features into an abstract vector space which has shown good performance, but requires much more computational power than HMM ba...

Citation Context

...the keyword is smaller than the distance D(x t , x − ) assigned to an utterance x − without the keyword. As recently proposed in our work on discriminative learning for phoneme-based keyword spotters =-=[15]-=-, a loss function suitable for the maximization of AUCk can be defined as, Lk = 1 |Rk||Rk| X x + ∈ Rk x − ∈ Rk lk(x + , x − ), where lk(x + , x − ) = max(0, 1 − D(x t , x − ) + D(x t , x + )). One can...

Citation Context

...l and learning algorithm, and describe experiments on detection of single-word terms in utterances from the Switchboard corpus. Our approach builds on previous work on discriminative keyword spotting =-=[2]-=-, adapting both the learning algorithm to handle unknown segmental time alignments, and feature functions to be similar to those used in tandem-based HMM systems. Adapting the algorithm to handle unkn...

Citation Context

...[3]; and hybrid KWS that combine phonetic and ASR-based KWS approaches to produce sub-word lattices for generating keyword search indices. A detailed survey of existing KWS techniques can be found in =-=[4, 5]-=-. In this work we focus on KWS in conversational speech that is distorted by the transmission channel with the resulting signal-to-noise ratios (SNR) ranging from 20 db to as low as 0 db collected und...

Citation Context

... synchronized “bundle”, as do all tongue features and the pair (glottis, velum), so K = 3. We now turn to constructing a spoken term detector and training it using a discriminative algorithm based on =-=[26]-=-. Our goal is to learn a function f : X ∗ × V ∗ → R, which takes as its input a speech utterance x ∈ X ∗ and a query term v ∈ V ∗ , where V is the vocabulary of words, and returns a score f(x, v) ∈ R ...

Citation Context

...ed on discrete HMMs were introduced [67]. Finally, discrete HMMs have been replaced by continuous HMMs [108]. There have been many attempts to do discriminative training, but the most relevant one is =-=[51]-=-, which targets a figure of merit instead of the error rate, and which is the baseline that I used in Section 6.2. 2.5 Computer Vision and Object Classification There has been a trend in object, acous...