Transcription of New Speaking Styles - Voicemail

In this paper we describe a new testbed for developing speech recognition algorithms a VoiceMail transcription task, analogous to other tasks such as the Switchboard, CallHome [1] and the Hub 4 tasks [2] which are currently used by speech recognition researchers. Spontaneous speech occurring in day-today life can broadly be classi ed into two categories (i) where the speaker does not receive any external feedback to direct his/her speech, and (ii) where the speaker receives external feedback from another person/machine/audience. Examples of the former category are radio broadcast news, voicemail etc., and examples of the latter category are telephone conversations, natural language transaction systems (eg. ATIS), seminars, etc. In general to obtain the best performance in transcribing a certain style of speech, it is necessary to train the speech recognition system on similar style of training data. Some of the speech categories mentioned above are quite well represented in currently existing databases. However, voicemail data is not well represented in any database, even though it represents a very large volume of real-world speech data. Consequently there is a need for a Voicemail database in order to improve transcription performance on a voicemail transcription task, and also to establish a new test bed for speech recognition algorithms. Similar to the Switchboard/CallHome databases, the Voicemail database comprises telephone bandwidth spontaneous speech. However the di erence with respect to the Switchboard and CallHome tasks is that the interaction is not between two humans, but rather between a human and a machine. Consequently, the speech is expected to be a little more formal in its nature, without the problems of cross-talk, barge-in etc. This eliminates some of the variables and provides more controlled conditions enabling one to concentrate on the aspects of spontaneous speech and e ects of the telephone channel. In this paper, we will describe the modality of collection of the speech data, and some algorithmic techniques that were devised based on this data. We will also describe the initial results of transcription performance on this task.