To obtain a robust acoustic model for a certain speech recognition task, a large amount of speech data is necessary. However, the preparation of speech data including recording and transcription is very costly and time-consuming. Although there are attempts to build generic acoustic models which are portable among different applications, speech recognition performance is typically task-dependent. This paper introduces a method for automatically building task-dependent acoustic models based on selective training. Instead of setting up a new database, only a small amount of task-specific development data needs to be collected. Based on the likelihood of the target model parameters given this development data, utterances which are acoustically close to the development data are selected from existing speech data resources. Since there are too many possibilities for selecting a data subset from a larger database in general, a heuristic has to be employed. The proposed algorithm deletes single utterances temporarily or alternates between successive deletion and addition of multiple utterances. In order to make selective training computationally practical, model retraining and likelihood calculation need to be fast. It is shown, that the model likelihood can be calculated fast and easily based on sufficient statistics without the need for explicit reconstruction of model parameters. The algorithm is applied to obtain an infant- and elderly-dependent acoustic model with only very few development data available. There is an improvement in word accuracy of up to 9% in comparison to conventional EM training without selection. Furthermore, the approach was also better than MLLR and MAP adaptation with the development data.

In this paper, the development, long-term operation and portability of a practical ASR application in a real environment is investigated. The target application is a speech-oriented guidance system installed at the local community center. The system has been exposed to ordinary people since November 2002. More than 300 hours or more than 700,000 inputs have been collected during four years. The outcome is a rare example of a large scale real-environment speech database. A simulation experiment is carried out with this database to investigate how the system's performance improves during the first two years of operation. The purpose is to determine empirically the amount of real-environment data which has to be prepared to build a system with reasonable speech recognition performance and response accuracy. Furthermore, the relative importance of developing the main system components, i.e. speech recognizer and the response generation module, is assessed. Although depending on the system's modeling capacities and domain complexity, experimental results show that overall performance stagnates after employing about 10-15 k utterances for training the acoustic model, 40-50 k utterances for training the language model and 40 k-50 k utterances for compiling the question and answer database. The Q&A database was most important for improving the system's response accuracy. Finally, the portability of the well-trained first system prototype for a different environment, a local subway station, is investigated. Since collection and preparation of large amounts of real data is impractical in general, only one month of data from the new environment is employed for system adaptation. While the speech recognition component of the first prototype has a high degree of portability, the response accuracy is lower than in the first environment. The main reason is a domain difference between the two systems, since they are installed in different environments. This implicates that it is imperative to take the behavior of users under real conditions into account to build a system with high user satisfaction.

Development of an ASR application such as a speech-oriented guidance system for a real environment is expensive. Most of the costs are due to human labeling of newly collected speech data to construct the acoustic model for speech recognition. Employment of existing models or sharing models across multiple applications is often difficult, because the characteristics of speech depend on various factors such as possible users, their speaking style and the acoustic environment. Therefore, this paper proposes a combination of unsupervised learning and selective training to reduce the development costs. The employment of unsupervised learning alone is problematic due to the task-dependency of speech recognition and because automatic transcription of speech is error-prone. A theoretically well-defined approach to automatic selection of high quality and task-specific speech data from an unlabeled data pool is presented. Only those unlabeled data which increase the model likelihood given the labeled data are employed for unsupervised training. The effectivity of the proposed method is investigated with a simulation experiment to construct adult and child acoustic models for a speech-oriented guidance system. A completely human-labeled database which contains real-environment data collected over two years is available for the development simulation. It is shown experimentally that the employment of selective training alleviates the problems of unsupervised learning, i.e. it is possible to select speech utterances of a certain speaker group but discard noise inputs and utterances with lower recognition accuracy. The simulation experiment is carried out for several selected combinations of data collection and human transcription period. It is found empirically that the proposed method is especially effective if only relatively few of the collected data can be labeled and transcribed by humans.