Hazen, T. J. and Zue, V. W. (1994). Automatic language identification using a segment-based approach. In Proceedings of the 3rd European Conference on Speech Communication and Technology, pages 1307--1310.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....procedure on independent material. Some researchers attempting to develop systems for automatic language identification (LID) have found F 0 to be of limited use in improving overall identification performance, but currently, LID is best performed using segmental and phonotactic information (Hazen and Zue, 1994). The most direct study of the utility of prosodic features in LID is summarized in Thym eGobbel and Hutchins (1996) Discrimination between pairs of English, Japanese, Mandarin and Spanish samples was attempted using 224 individual features and feature pairs. F 0 based features included several ....

....to a series of prompts designed to elicit speech on a continuum from specific lexical items to unconstrained monologue. The database includes approximately 200 speakers for each of 11 languages. Several groups have published results based on this corpus, making comparison across systems feasible (Hazen and Zue, 1994; Zissman, 1995; Thym e Gobbel and Hutchins, 1996) The OGI corpus provides recordings acquired over commercial telephone lines of speakers responses to an automatically generated series of prompts. The prompts elicit lexically constrained responses (e.g. Please recite the seven days of the ....

Hazen, T. J. and Zue, V. W. (1994). Automatic language identification using a segment-based approach. In Proceedings of the 3rd European Conference on Speech Communication and Technology, pages 1307--1310.

.... classifiers[2] vector quantization[3, 9, 28] or neural nets[20] Broad phonetic labels were used with finite state models[15] and with neural nets[20] More recently, Gaussian mixture and HMM have been proposed for language identification[21, 31] as well as stochastic segment based models[10]. This paper presents our recent work in language identification using phone based acoustic likelihoods[5, 13] The basic idea is to process in parallel the unknown incoming speech by different sets of phone models (each set is a large ergodic HMM) for each of the languages under consideration, ....

....Our approach for French English identification took advantage of the associated phonetic transcriptions, whereas for the OGI data, training was performed without the use of transcriptions. Despite these conditions, our results compare favorably to previously published results on the same corpus[20, 31, 10]. ....

T.J. Hazan, V.W. Zue, "Automatic Language Identification using a Segment-Based Approach," EUROSPEECH-93.

....Modeling of Sequences. 2. Sequence Selection and Error Estimation. 3. Phoneme Merging. The flowchart shown in Figure 3.7 depicts the iterative process of merging phonemes and estimating language classification error. At the extremes of this algorithm we obtain either a phoneme based system [36] or a broad category based system [76] This algorithm will return the lowest number of phoneme clusters without losing language discriminability, together with an error estimate as a function of time due to the chosen features. The input to this algorithm will be both the labeled and the ....

Hazen, T. J. Automatic Language Identification Using a Segment-based Approach. Master's thesis, Massachusetts Institute of Technology, Aug. 1993.

....As a result, the review does not contain any mention of the author s published work on automatic language identification using high quality speech [MC92a] or telephone speech [MC92b] These are described in chapters 4 and 5 respectively. Marc Zissman s work [Zis93] and Hazen and Zue s work [HZ93] using the OGI Multi language Telephone Speech Corpus ( MCO92] and Chapter 3) are described in Chapter 6. Also, since the focus of this review is on studies in automatic language identification, Atkinson s work on human listening experiments with English and Spanish [Atk68] is excluded. The ....

....0 and L Gamma Other tasks were also comparable to our results. He did not examine the English GammaL 0 Gamma Other task. Timothy Hazen and Victor Zue at MIT have reported results on a segment based approach to automatic language identification, designed around a formal probabilistic framework [HZ93] Using probablistic models for the phonotactic, prosodic and acoustic properties of the different languages in the corpus, they obtained an identification accuracy of 47.7 on the development test set for the ten language task. Results on the English GammaL 0 , L Gamma Other and ....

T. J. Hazen and V. W. Zue. Automatic language identification using a segment-based approach. In Proceedings 3rd European Conference on Speech Communication and Technology (Eurospeech 93), Berlin, Germany, September 1993.

....to a series of prompts designed to elicit speech on a continuum from specific lexical items to unconstrained monologue. The database includes approximately 200 speakers for each of 11 languages. Several groups have published results based on this corpus, making comparison across systems feasible (Hazen and Zue, 1994; Zissman, 1995; Thym e Gobbel and Hutchins, 1996) LID systems typically make use of several more or less independent sources of information which are extracted from the raw speech signal (Hazen and Zue, 1997) By far the most important source of information for LID is the short term spectrum ....

Hazen, T. J. and Zue, V. W. (1994). Automatic language identification using a segment-based approach.

....Laboratories, Natural Speech Technologies, OGI, and RPI) are participating in this ongoing evaluation. In addition, the last two years have seen a substantial increase in papers on language ID in major speech conferences and symposia such as ICASSP, Eurospeech and SRS (Speech Research Symposium) [7, 15, 20, 23, 24, 28, 33, 34], with complete sessions devoted to language ID in each of them. This proliferation of different approaches to the problem using the same corpus has led to an open exchange of ideas a process so essential for research progress. 5 Perceptual Studies on Language ID 5.1 Perceptual Benchmarks are ....

....span several phonemes) should contribute much to speech recognition. This insight has, however, not contributed much to the success of current systems. Similarly, the incorporation of explicit prosodic information was not as useful in early language identification systems of the current generation[7, 22] as the designers may have hoped. There is, nonetheless, much reason to think that prosodic differences will contribute significantly to language identification in the future, and recent research has begun to fulfill this promise[10] Hazen and Zue [7] incorporated pitch information by ....

[Article contains additional citation context not shown here]

T. J. Hazen and V. W. Zue. Automatic language identification using a segment-based approach. In Proceedings 3rd European Conference on Speech Communication and Technology (Eurospeech 93), September 1993.

.... Hidden Markov Models (HMM) UN90, SAG91, RMM91, NUS92, Zis93] neural networks [MC92] embedded keywords [RR94] syllabic spectral features[Li94] the distinction between polyphonemes and monophonemes [ADB94] and subword recognition exploiting acoustic phonotactic constraints [MBA 93, TCP94, HZ93, HZ94, RSN94, LG94, KH94, ZS94, BABC94] Systematic reviews of language identification activities can be found in[Mut93, MBC94] These efforts have improved the performance of LID systems dramatically. It was shown that accurate classification could not be achieved by simply using frame based ....

....identification by other 14 Recognizer 1 Recognizer 2 Recognizer M LID model 1,2, N Score Generator 1 LID model 1,2, N LID model 1,2, N Score Generator 2 Score Generator M LID Result Speech Signal Final Classifier Figure 1. 3: The baseline system researchers (e.g. MBA 93, ZS94, HZ93] etc) In this dissertation, this idea is extended. 1.3.3 Contributions: Methods Proposed to Improve the Baseline System The major contributions of this dissertation are: 1. Increasing phonotactic modeling accuracy without drastically increasing the amount of required training data. A language ....

[Article contains additional citation context not shown here]

T.J. Hazen and V.W. Zue. Automatic language identification using a segment-based approach. In Proceedings Eurospeech 93, pages 1303--1306, September 1993.

....is intended to provide a fast match list of likely candidate 1 This research was supported by ARPA under Contract N001489 J 1332 monitored through the Office of Naval Research and by a grant from Texas Instruments. languages for a spoken utterance. This system, previously described in [1] and [2], incorporates phonetic, acoustic and prosodic information within a probabilistic framework. SYSTEM DESCRIPTION Corpus The ALI system described herein was trained and tested using the OGI Multi Language Telephone Speech Corpus [3] The original OGI database consisted of utterances spoken in 10 ....

....is achieved. RESULTS Development Results As shown in the first row of Table 1, our first implementation of an ALI system, as reported in [1] achieved an accuracy of 47.7 when tested on the complete de Date Comments Accuracy 4 93 System presented in [1] 47.7 8 93 System presented in [2] 48.6 1 94 Channel normalization 54.8 1 94 Mixture Gaussian duration model 55.8 2 94 Recognizer trained w OGI data 58.5 Table 1: Summary of development test results Test date March 94 June 94 Utt. length 10 sec. 30 sec. 10 sec. 30 sec. Language model 61.6 72.7 62.7 77.5 ....

[Article contains additional citation context not shown here]

T. J. Hazen, Automatic Language Identification Using a Segment-Based Approach, SM thesis, MIT, 1993.

....which is intended to provide a fast match list of likely candidate 1 This research was supported by ARPA under Contract N001489 J 1332 monitored through the Office of Naval Research and by a grant from Texas Instruments. languages for a spoken utterance. This system, previously described in [1] and [2] incorporates phonetic, acoustic and prosodic information within a probabilistic framework. SYSTEM DESCRIPTION Corpus The ALI system described herein was trained and tested using the OGI Multi Language Telephone Speech Corpus [3] The original OGI database consisted of utterances ....

....hill climbing search routine where the system s scaling factors are optimized one model at a time in an iterative fashion until a local maximum in performance is achieved. RESULTS Development Results As shown in the first row of Table 1, our first implementation of an ALI system, as reported in [1], achieved an accuracy of 47.7 when tested on the complete de Date Comments Accuracy 4 93 System presented in [1] 47.7 8 93 System presented in [2] 48.6 1 94 Channel normalization 54.8 1 94 Mixture Gaussian duration model 55.8 2 94 Recognizer trained w OGI data 58.5 Table 1: ....

[Article contains additional citation context not shown here]

T. J. Hazen and V. W. Zue, "Automatic language identification using a segment-based approach," In Proc. Eurospeech 93, pp. 1303-1306, 1993.

No context found.

T. J. Hazen and V. W. Zue. Automatic language identification using a segmentbased approach. In Proceedings of the 3rd European Conference on Speech Communication and Technology, pages 1307--1310, 1993.

No context found.

T. J. Hazen. Automatic language identification using a segment-based approach. Master's thesis, Massachusetts Institute of Technology, August 1993.

....of Naval Research and by a grant from Texas Instruments. merits of each model as system parameters are varied, we will nevertheless measure overall system performance using a publicly available multi language corpus. Interested readers are referred to a more detailed description of this work in [10]. PROBABILISTIC FRAMEWORK Before designing the segment based ALI system, a probabilistic framework describing the ALI problem was derived. To begin, let L = fL 1 ; L 2 ; L n g represent the language set of n different languages. An ALI system s basic objective is to determine which of ....

T. J. Hazen, Automatic Language Identification Using a Segment-Based Approach, SM thesis, MIT, 1993.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC