Large Vocabulary Continuous Speech Recognition 3
HomeFull List of Titles
1: ICSLP'98 Proceedings
Keynote SpeechesText-To-Speech Synthesis 1Spoken Language Models and Dialog 1Prosody and Emotion 1Hidden Markov Model Techniques 1Speaker and Language Recognition 1Multimodal Spoken Language Processing 1Isolated Word RecognitionRobust Speech Processing in Adverse Environments 1Spoken Language Models and Dialog 2Articulatory Modelling 1Talking to Infants, Pets and LoversRobust Speech Processing in Adverse Environments 2Spoken Language Models and Dialog 3Speech Coding 1Articulatory Modelling 2Prosody and Emotion 2Neural Networks, Fuzzy and Evolutionary Methods 1Utterance Verification and Word Spotting 1 / Speaker Adaptation 1Text-To-Speech Synthesis 2Spoken Language Models and Dialog 4Human Speech Perception 1Robust Speech Processing in Adverse Environments 3Speech and Hearing Disorders 1Prosody and Emotion 3Spoken Language Understanding Systems 1Signal Processing and Speech Analysis 1Spoken Language Generation and Translation 1Spoken Language Models and Dialog 5Segmentation, Labelling and Speech Corpora 1Multimodal Spoken Language Processing 2Prosody and Emotion 4Neural Networks, Fuzzy and Evolutionary Methods 2Large Vocabulary Continuous Speech Recognition 1Speaker and Language Recognition 2Signal Processing and Speech Analysis 2Prosody and Emotion 5Robust Speech Processing in Adverse Environments 4Segmentation, Labelling and Speech Corpora 2Speech Technology Applications and Human-Machine Interface 1Large Vocabulary Continuous Speech Recognition 2Text-To-Speech Synthesis 3Language Acquisition 1Acoustic Phonetics 1Speaker Adaptation 2Speech Coding 2Hidden Markov Model Techniques 2Multilingual Perception and Recognition 1Large Vocabulary Continuous Speech Recognition 3Articulatory Modelling 3Language Acquisition 2Speaker and Language Recognition 3Text-To-Speech Synthesis 4Spoken Language Understanding Systems 4Human Speech Perception 2Large Vocabulary Continuous Speech Recognition 4Spoken Language Understanding Systems 2Signal Processing and Speech Analysis 3Human Speech Perception 3Speaker Adaptation 3Spoken Language Understanding Systems 3Multimodal Spoken Language Processing 3Acoustic Phonetics 2Large Vocabulary Continuous Speech Recognition 5Speech Coding 3Language Acquisition 3 / Multilingual Perception and Recognition 2Segmentation, Labelling and Speech Corpora 3Text-To-Speech Synthesis 5Spoken Language Generation and Translation 2Human Speech Perception 4Robust Speech Processing in Adverse Environments 5Text-To-Speech Synthesis 6Speech Technology Applications and Human-Machine Interface 2Prosody and Emotion 6Hidden Markov Model Techniques 3Speech and Hearing Disorders 2 / Speech Processing for the Speech and Hearing Impaired 1Human Speech ProductionSegmentation, Labelling and Speech Corpora 4Speaker and Language Recognition 4Speech Technology Applications and Human-Machine Interface 3Utterance Verification and Word Spotting 2Large Vocabulary Continuous Speech Recognition 6Neural Networks, Fuzzy and Evolutionary Methods 3Speech Processing for the Speech-Impaired and Hearing-Impaired 2Prosody and Emotion 72: SST Student Day
SST Student Day - Poster Session 1SST Student Day - Poster Session 2Author Index
A B C D E F G H I
J K L M N O P Q R
S T U V W X Y Z
Multimedia Files
The BBN Single-Phonetic-Tree Fast-Match Algorithm
Authors:
Long Nguyen, BBN Technologies, GTE Internetworking (USA)
Richard Schwartz, BBN Technologies, GTE Internetworking (USA)
Page (NA) Paper number 1028
Abstract:
In January 1993, BBN demonstrated for the first time a real-time, 20K-word, speaker-independent, continuous speech recognition system, implemented in software on an off-the-shelf workstation. The key to the real-time system was a novel, proprietary fast-match algorithm which had two important properties: high-accuracy recognition and run-time proportional to only the cube root of the vocabulary size. This paper describes that fast-match algorithm in detail. While a number of fast-match algorithms have been published, the BBN algorithm continues to have novel features that have not appeared in the literature. In this fast-match, the vocabulary is organized as a phonetic tree. The last phonemes of the words always locate at the leaves. Each node of the phonetic tree is assigned a set id representing a group of words which share this node. The acoustic models associated with the nodes are the composite triphones where there could be more than one right context. The language models used in this fast-match are the Pr(set_of_words|some_word) which are evaluated at every node of the words during the search. The search itself is similar to the usual beam search with one addition: To activate a node, we temporarily use some set bigrams; to leave that node, we take out that temporary bigram.
SL981028.PDF (From Author)
SL981028.PDF (Rasterized)
TOP
An Efficient Two-pass Search Algorithm Using Word Trellis Index
Authors:
Akinobu Lee, School of Informatics, Kyoto University (Japan)
Tatsuya Kawahara, School of Informatics, Kyoto University (Japan)
Shuji Doshita, School of Informatics, Kyoto University (Japan)
Page (NA) Paper number 655
Abstract:
We propose an efficient two-pass search algorithm for LVCSR. Instead of conventional word graph, the first preliminary pass generates "word trellis index", keeping track of all survived word hypotheses within the beam every time-frame. As it represents all found word boundaries non-deterministically, we can (1) obtain accurate sentence-dependent hypotheses on the second search, and (2) avoid expensive word-pair approximation on the first pass. The second pass performs an efficient stack decoding search, where the index is referred to as predicted word list and heuristics. Experimental results on 5,000-word Japanese dictation task show that, compared with the word-graph method, this trellis-based method runs with less than 1/10 memory cost while keeping high accuracy. Finally, by handling inter-word context dependency, we achieved the word error rate of 5.6%.
SL980655.PDF (From Author)
SL980655.PDF (Rasterized)
TOP
Nozomi -- a Fast, Memory-Efficient Stack Decoder For LVCSR
Authors:
Mike Schuster, ATR, Interpreting Telecommunications Laboratories (Japan)
Page (NA) Paper number 464
Abstract:
This paper describes some of the implementation details of the ``Nozomi'' stack decoder for LVCSR. The decoder was tested on a Japanese Newspaper Dictation Task using a 5000 word vocabulary. Using continuous density acoustic models with 2000 and 3000 states trained on the JNAS/ASJ corpora and a 3-gram LM trained on the RWC text corpus, both models provided by the IPA group, it was possible to reach more than 95% word accuracy on the standard test set. With computationally cheap acoustic models we could achieve around 89% accuracy in nearly realtime on a 300 Mhz Pentium II. Using a disk-based LM the memory usage could be optimized to 4 MB in total.
SL980464.PDF (From Author)
SL980464.PDF (Rasterized)
TOP
Reducing the OOV Rate in Broadcast News Speech Recognition
Authors:
Thomas Kemp, ISL, University of Karlsruhe (Germany)
Alex Waibel, ISL, University of Karlsruhe (Germany)
Page (NA) Paper number 757
Abstract:
To achieve the long-term goal of robust, real-time broadcast news transcription, several problems have to be overcome, e.g. the variety of acoustic conditions and the unlimited vocabulary. In this paper we address the problem of unlimited vocabulary. We show, that this problem is more serious for German than it is for English. Using a speech recognition system with a large vocabulary, we dynamically adapt the active vocabulary to the topic of the current news segment. This is done by using information retrieval (IR) techniques on a large collection of texts automatically gathered from the internet. The same technique is also used to adapt the language model of the recognition system. The process of vocabulary adaptation and language model retraining is completely unsupervised. We show, that dynamic vocabulary adaptation can significantly reduce the out-of-vocabulary (OOV) rate and the word error rate of our broadcast news transcription system View4You.
SL980757.PDF (From Author)
SL980757.PDF (Rasterized)
TOP
Using Automatically-Derived Acoustic Sub-word Units in Large Vocabulary Speech Recognition
Authors:
Michiel Bacchiani, Boston University, ECE Department (USA)
Mari Ostendorf, Boston University, ECE Department (USA)
Page (NA) Paper number 586
Abstract:
Although most parameters in a speech recognition system are estimated from data, the unit inventory and lexicon are generally hand crafted and therefore unlikely to be optimal. This paper describes a joint solution to the problems of learning a unit inventory and corresponding lexicon from data. The methodology, which requires multiple training tokens per word, is then extended to handle infrequently observed words using a hybrid system that combines automatically-derived units with phone-based units. The hybrid system outperforms a phone-based system in first-pass decoding experiments on a large vocabulary conversational speech recognition task.
SL980586.PDF (From Author)
SL980586.PDF (Rasterized)
TOP
Fabricating Conversational Speech Data with Acoustic Models: a Program to Examine Model-Data Mismatch
Authors:
Don McAllaster, Dragon Systems (USA)
Lawrence Gillick, Dragon Systems (USA)
Francesco Scattone, Dragon Systems (USA)
Michael Newman, Dragon Systems (USA)
Page (NA) Paper number 986
Abstract:
We present a study of data simulated using acoustic models trained on Switchboard data, and then recognized using various Switchboard-trained models. Simple development models give a word error rate (WER) of about 47%, when recognizing real Switchboard conversations. If we simulate speech from word transcriptions, obtaining the word pronunciations from our recognition dictionary, the WER drops by a factor of five to ten. If we use more realistic hand-labeled phonetic transcripts to fabricate data, we obtain WERs in the low 40's, close to those found in actual speech data. These and other experiments we describe in the paper suggest that there is a substantial mismatch between real speech and the combination of our acoustic models and the pronunciations in our recognition dictionary. The use of simulation in speech recognition research appears to be a promising tool in our efforts to understand and reduce the size of this mismatch.