Phoneme Based Acoustics Keyword Spotting in Informal Continuous Speech

The Spoken Web Search Task at MediaEval 2011

“…Following [19], an HMM was built for each query using a dictionary, and the log likelihood ratio between a query model and a background model (free phone loop) was computed. The development forced alignment and graphemic transcription of queries to obtain reference pronunciation of each query were used.…”

Improving acoustic based keyword spotting using LVCSR lattices

“…Parallelly concatenated keyword models are then accompanied by filler and background models (represented by simple phone loops) to create a decoding network, as shown in Fig 1. Likelihoods of the detected keywords are taken from the last state of each keyword model (computed using Viterbi decoder) and compared with the likelihood obtained from the background model. Confidence Score (CS) of each detected keyword is then given as a log-likelihood ratio between these two likelihoods [1]. Such the acoustic KWS is denoted as KW S 1xRT acoust and is able to run much faster than LVCSR-KWS (far below 1xRT).…”

“…Unlike ASR, the acoustic KWS does not need to recognize the whole sentence. The keywords are searched in parameterized spoken data (acoustic features) [1]. Unlike acoustic KWS, Large Vocabulary Continuous Speech Recognition (LVCSR) based KWS systems (often called Spoken Term Detection systems) search keywords in the output of the LVCSR, i.e., word recognition strings -lattices [2].…”

“…The acoustic KWS described above represents a second potential approach towards open-vocabulary KWS, where a model of any keyword (composed of acoustic phone models) is built at the time the keyword is entered [1]. Although this acoustic KWS approach has not been thoroughly investigated and usually performs worse than LVCSR based techniques, its simplicity, speed and robustness to OOVs can be of interest for real-time applications [3].…”

Point process models of spectro-temporal modulation events for speech recognition

“…The first was a keyword-filler hidden Markov modelbased system with an architecture similar to [15], built on top of a Gaussian mixture model (GMM) based monophone acoustic model (one 8 component GMM per phone with a standard 39-dimensional mel frequency cepstral coefficient observation space, trained on the TIMIT sx/i sentences). Each keyword HMM consisted of one pronunciation path per word with the phone GMMs defining the emission densities.…”