Combined low level and high level features for Out-Of- Vocabulary Word detection - Archive ouverte HAL Accéder directement au contenu
Poster De Conférence Année : 2009

Combined low level and high level features for Out-Of- Vocabulary Word detection

Résumé

Principle-Extract low and high level features (acoustic, search graph topology, linguistic) for each word-Features are classified with a boosting algorithm-A semantic modul refines the detection process Acoustic features :-Log-likelihood-Average log-likelihood per frame-Difference between the word log-likelihood and the unconstrained acoustic decoding of the corresponding speech segment. Linguistic features :-3-gram probabilities-perplexity of the word in the window-unigram probability-Current backoff level of the word Graph features :-Based on word confusion network-posteriors-number of alternative paths-distribution of posteriors (min, max, mean)-number of null links in the window (500 ms) Word 1 Word 3 Word 2 Boosting algorithm OOV Not OOV Semantic module OOV Not OOV-Input vector composed by 3 consecutive words-Each word is composed of 23 features-The final vector is 69 coefficiens The semantic module Combination of two measures :-WEB : estimates the probability of word co-occurrences as a ratio of Google hits.-Gigaword Corpus : based on Latent Semantic Analyis (LSA), to estimate how much the targeted word is semantically close to the current segment. Experimental framework Semantic filtering and Conclusions The Lia broacast news system used in ESTER campaign-HMM-based decoder Speeral-Asynchronous decoder operating on a phoneme lattice-Acoustics models are HMM-based with cross word triphones-Language model is 3-gram estimated on 200M words-Lexicon is 67 Kwords-One pass in 3RT The ESTER Corpus :-French radio broadcasts-Training for OOV provided by ESTER-2 train (100 hours)-15K OOV and 1 M of words-The test is the ESTER test : 7 hours → 982 OOVS for 70011 word (1.04% OOV)-OOV test overlapping with OOV train is 2% Detection protocol : OOV words have been manually specified by selecting all the reference words not available in the lexicon. During the detection, if a marked OOV word overlaps with a true OOV word, the true OOV word is considered as detected. In all other cases we consider a marked word as a false detection. Semantic filtering Semantic module is used only on detected OOV words, with the purpose to refine detection on semantically coherent words. The ROC curves are presented in Figure 5. Results show that the filter reduces the EER by 4% relatively (15.2% to 14.6%), while the false detection rate decreases by about 5% relative. Conclusions Our experiments showed promising results: OOV word detection seems to be homogeneous, despite the inconstant WER. The proposed method allows one to detect 43% of the unknown words with a 2.5% false acceptance rate, or 90% for 17.5% false acceptance. Experiments show that linguistic and graph-based features are the most relevant predictors. However, acoustic features associated to the others make the detection more robust. Finally, semantic filtering provides a slight but significant improvement.
Fichier principal
Vignette du fichier
posterInterspeech2009.pdf (1.09 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-02088875 , version 1 (03-04-2019)

Identifiants

  • HAL Id : hal-02088875 , version 1

Citer

Benjamin Lecouteux, Georges Linarès, Benoit Favre. Combined low level and high level features for Out-Of- Vocabulary Word detection. Interspeech, 2009, Brighton, United Kingdom. ⟨hal-02088875⟩
65 Consultations
23 Téléchargements

Partager

Gmail Facebook X LinkedIn More