Detection of unseen words in conversational Mandarin

Bufyko, Ivan; Kimball, Owen; Siu, Man-Hung; Herrero, Jose L.; Blum, Dan

doi:10.1109/icassp.2012.6289087

Cited by 19 publications

(11 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this section we introduce a novel pFA-based score normalization method, where we estimate the false alarm rate for a given keyword at various levels of the raw score (log probability of the DP alignment between the query and the recognition output represented by a phonetic consensus network -see [3] for details). We model the mapping from the raw log probability to the log of FA rate with a straight line, computed for each keyword separately.…”

Section: Unsupervised Score Normalization Using Linear Fitmentioning

confidence: 99%

See 1 more Smart Citation

Normalization of phonetic keyword search scores

Karakos

Bulyko

Schwartz

et al. 2014

2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

As shown in [1,2], score normalization is of crucial importance for improving the Average Term-Weighted Value (ATWV) measure that is commonly used for evaluating keyword spotting systems. In this paper, we compare three different methods for score normalization within a keyword spotting system that employs phonetic search. We show that a new unsupervised linear fit method results in better-estimated posterior scores, that, when fed into the keyword-specific normalization of [1], result in ATWV gains of 3% on average. Furthermore, when these scores are used as features within a supervised machine learning framework, they result in additional gains of 3.8% on average over the five languages used in the first year of the IARPA-funded project Babel.

show abstract

Section: Unsupervised Score Normalization Using Linear Fitmentioning

confidence: 99%

“…Of course, instead of phones, one can choose to use some other representation that will achieve the goal of covering a large class of OOV words. The procedure for doing the phonetic search in this paper follows closely the one presented in [3].…”

Section: Introductionmentioning

confidence: 97%

Normalization of phonetic keyword search scores

Karakos

Bulyko

Schwartz

et al. 2014

2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

show abstract

“…A hybrid recognizer with both word and subword units is used to generate lattices containing [6], or the lattices generated by a word recognizer and a subword recognizer are combined for the detection of OOV words [7]. A mix two-stage LM also has been proposed for OOV words recognition [8].…”

Section: Introductionmentioning

confidence: 99%

Enhanced Out of Vocabulary Word Detection Using Local Acoustic Information

Wang

Zhang

et al. 2014

2014 Tenth International Conference on Intelligent Information Hiding and Multimedia Signal Processing

View full text Add to dashboard Cite

The detection of Out-of-vocabulary (OOV) words is a crucial problem for spoken term detection (STD). In this paper, the use of integration with local acoustic information is investigated to retrieve more OOV words. Tokens with high local acoustic probabilities propagated in the search space at the decoding stage will be forced to propagate to the next frame. In this way, acoustic similar words can be reserved in recognition results without considering of language model probabilities. Experimental results show that this new approach results in a significant increase in the performance of OOV words detection. At least a relative improvement of 8.5% in equal error rate is achieved over the baseline system. Meanwhile, it will do no harm to in-vocabulary (IV) words detection. With some refinement of beam pruning, the decoding time only rises 3% relative to the baseline system.

show abstract

“…Such techniques are often used in combination with LVCSR-based methods because their search accuracy for IV terms tends to be lower than that of LVCSR-based methods. Recently, methods that combine multiple types of indices were proposed that achieve high detection accuracy [8,9,10]. For example, the best performance in the latest NTCIR STD evaluation [2] was obtained by a method that combines 10 different recognizers' outputs [10].…”

Section: Introductionmentioning

confidence: 99%

Multiple index combination for Japanese spoken term detection with optimum index selection based on OOV-region classifier

Kanda

Itoyama

Okuno

2013

2013 IEEE International Conference on Acoustics, Speech and Signal Processing

View full text Add to dashboard Cite

In this paper, a novel index combination method for spoken term detection is proposed. In our method, outputs from four different recognizers (word, syllable, word-syllable, and fragment recognizer) are combined into one confusion network. A novel index-selection method for the multiple index-combination method is then used to suppress the increase of the index size. Two methods are proposed to reduce index size: (1) arc selection and (2) unit selection, both of which are based on an OOV-region classifier score. Experimental results with 39 hours of Japanese lecture recordings showed that the index-selection method achieved a 22% reduction of index size of the best confusion network while maintaining its high accuracy.Compared with the best phoneme-based index from a single recognizer, the proposed method achieved a 25.0% and 14.8% relative error reduction for IV and OOV queries without increasing the index size.

show abstract

Detection of unseen words in conversational Mandarin

Cited by 19 publications

References 13 publications

Normalization of phonetic keyword search scores

Normalization of phonetic keyword search scores

Enhanced Out of Vocabulary Word Detection Using Local Acoustic Information

Multiple index combination for Japanese spoken term detection with optimum index selection based on OOV-region classifier

Contact Info

Product

Resources

About