• MALSORI
• /
• v.55
• /
• pp.77-91
• /
• 2005

In this paper, we propose an XML based meta-data specification for industrial speech databases. Building speech databases is very time-consuming and expensive. Recently, by the government supports, huge amount of speech corpus has been collected as speech databases. However, the formats and meta-data for speech databases are different depending on the constructing institutions. In order to advance the reusability and portability of speech databases, a standard representation scheme should be adopted by all speech database construction institutions. ETRI proposed a XML based annotation scheme [51 for speech databases, but the scheme has too simple and flat modeling structure, and may cause duplicated information. In order to overcome such disadvantages in this previous scheme, we first define the speech database more formally and then identify object appearing in speech databases. We then design the data model for speech databases in an object-oriented way. Based on the designed data model, we develop the meta-data specification for industrial speech databases.

• Proceedings of the KSPS conference
• /
• 2003.05a
• /
• pp.131-134
• /
• 2003

This paper describes an efficient algorithm to generate compact and complete prompts lists for connected spoken digits database. In building a connected spoken digit recognizer, we have to acquire speech data in various contexts. However, in many speech databases the lists are made by using random generators. We provide an efficient algorithm that can generate compact and complete lists of digits in various contexts. This paper includes the proof of optimality and completeness of the algorithm.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.14 no.3
• /
• pp.357-362
• /
• 2004

Language models are essential in predicting the next word in a spoken sentence, thereby enhancing the speech recognition accuracy, among other things. However, spoken language domains are too numerous, and therefore developers suffer from the lack of corpora with sufficient sizes. This paper proposes a method of combining two n-gram language models, one constructed from a very small corpus of the right domain of interest, the other constructed from a large but less adequate corpus, resulting in a significantly enhanced language model. This method is based on the observation that a small corpus from the right domain has high quality n-grams but has serious sparseness problem, while a large corpus from a different domain has more n-gram statistics but incorrectly biased. With our approach, two n-gram statistics are combined by extending the idea of Katz's backoff and therefore is called a dual-source backoff. We ran experiments with 3-gram language models constructed from newspaper corpora of several million to tens of million words together with models from smaller broadcast news corpora. The target domain was broadcast news. We obtained significant improvement (30%) by incorporating a small corpus around one thirtieth size of the newspaper corpus.

• Phonetics and Speech Sciences
• /
• v.1 no.3
• /
• pp.49-55
• /
• 2009

This study investigates whether rhythm metrics can be used to capture the rhythmic differences between spontaneous and read English speech. Transcription of spontaneous speech tokens extracted from a corpus is read by three English native speakers to generate the corresponding read speech tokens. Two data sets are compared in terms of seven rhythm measures that are suggested by previous studies. Results show that there is a significant difference in the values of vowel-based metrics (VarcoV and nPVI-V) between spontaneous and read speech. This manifests a greater variability in vocalic intervals in spontaneous speech than in read speech. The current study is especially meaningful as it demonstrates a way in which speech styles can be differentiated and parameterized in numerical terms.

• Phonetics and Speech Sciences
• /
• v.8 no.2
• /
• pp.31-39
• /
• 2016

In emergency dispatching at 119 Command & Dispatch Center, some inconsistencies between the 'standard emergency aid system' and 'dispatch protocol,' which are both mandatory to follow, cause inefficiency in the dispatcher's performance. If an emergency dispatch system uses automatic speech recognition (ASR) to process the dispatcher's protocol speech during the case registration, it instantly extracts and provides the required information specified in the 'standard emergency aid system,' making the rescue command more efficient. For this purpose, we have developed a Korean large vocabulary continuous speech recognition system for 400,000 words to be used for the emergency dispatch system. The 400,000 words include vocabulary from news, SNS, blogs and emergency rescue domains. Acoustic model is constructed by using 1,300 hours of telephone call (8 kHz) speech, whereas language model is constructed by using 13 GB text corpus. From the transcribed corpus of 6,600 real telephone calls, call logs with emergency rescue command class and identified major symptom are extracted in connection with the rescue activity log and National Emergency Department Information System (NEDIS). ASR is applied to emergency dispatcher's repetition utterances about the patient information. Based on the Levenshtein distance between the ASR result and the template information, the emergency patient information is extracted. Experimental results show that 9.15% Word Error Rate of the speech recognition performance and 95.8% of emergency response detection performance are obtained for the emergency dispatch system.

• Journal of Computing Science and Engineering
• /
• v.2 no.2
• /
• pp.162-179
• /
• 2008

We have developed a tool for annotation of electronic health record (EHR) data. Currently we are in the process of manually annotating a corpus of Norwegian general practitioners' EHRs with mainly linguistic information. The purpose of this project is to attain a linguistically annotated corpus of patient histories from general practice. This corpus will be put to future use in medical language processing and information extraction applications. The paper outlines some of our practical experiences from developing such a corpus and, in particular, the effects of semi-automated annotation. We have also done some preliminary experiments with part-of-speech tagging based on our corpus. The results indicated that relevant training data from the clinical domain gives better results for the tagging task in this domain than training the tagger on a corpus form a more general domain. We are planning to expand the corpus annotations with medical information at a later stage.

• Speech Sciences
• /
• v.15 no.4
• /
• pp.97-105
• /
• 2008

Nowadays an HMM-based text-to-speech system (HTS) has been very widely studied because it needs less memory and low computation complexity and is suitable for embedded systems in comparison with a corpus-based unit concatenation text-to-speech one. It also has the advantage that voice characteristics and the speaking rate of the synthetic speech can be converted easily by modifying HMM parameters appropriately. We implemented an HMM-based Korean text-to-speech system using a small size Korean speech DB and proposes a method to increase the naturalness of the synthetic speech by controlling duration model parameters in the HMM-based Korean text-to speech system. We performed a paired comparison test to verify that theses techniques are effective. The test result with the preference scores of 73.8% has shown the improvement of the naturalness of the synthetic speech through controlling the duration model parameters.

• MALSORI
• /
• no.47
• /
• pp.109-120
• /
• 2003

This paper presents about the activities of speech database standardization in ETRI. Recently, with the support of government, ETRI and SiTEC have been gathering the large speech corpus for the domestic speech related companies. First, due to the lack of sharing the knowledge of speech database specification, the distributed speech database has a different format. Hence it seems to be needed to have the same format as soon as possible. ETRI and SiTEC are trying to find the better representation format of speech database. Second, we introduce a new description method of the annotation information of speech database. As one of the structured description method, XML based description will be applied to represent the metadata of the speech database. It will be continuously revised through the speech technology standard forum during this year.

• ETRI Journal
• /
• v.35 no.1
• /
• pp.100-108
• /
• 2013

In this paper, a feature extraction (FE) method is proposed that is comparable to the traditional FE methods used in automatic speech recognition systems. Unlike the conventional spectral-based FE methods, the proposed method evaluates the similarities between an embedded speech signal and a set of predefined speech attractor models in the reconstructed phase space (RPS) domain. In the first step, a set of Gaussian mixture models is trained to represent the speech attractors in the RPS. Next, for a new input speech frame, a posterior-probability-based feature vector is evaluated, which represents the similarity between the embedded frame and the learned speech attractors. We conduct experiments for a speech recognition task utilizing a toolkit based on hidden Markov models, over FARSDAT, a well-known Persian speech corpus. Through the proposed FE method, we gain 3.11% absolute phoneme error rate improvement in comparison to the baseline system, which exploits the mel-frequency cepstral coefficient FE method.

• Phonetics and Speech Sciences
• /
• v.3 no.3
• /
• pp.115-121
• /
• 2011

This paper examines how L2 speech production varies according to learner's L2 proficiency level. L2 speech production variations are analyzed by quantitative measures at word and phone levels using Korean learners' English corpus. Word-level variations are analyzed using correctness to explain how speech realizations are different from the canonical forms, while accuracy is used for analysis at phone level to reflect phone insertions and deletions together with substitutions. The results show that speech production of learners with different L2 proficiency levels are considerably different in terms of performance and individual realizations at word and phone levels. These results confirm that speech production of non-native speakers varies according to their L2 proficiency levels, even though they share the same L1 background. Furthermore, they will contribute to improve non-native speech recognition performance of ASR-based English language educational system for Korean learners of English.