The telephone network presents speech recognition devices with a band-limited, noisy, and, in some cases, distorted speech signal. A series of experiments were performed to quantify the effects of these transformations on two current recognition algorithms: a) an acoustic segmentation algorithm and b) an acoustic classification algorithm. The data used in these experiments are a subset of the TIMIT speech database and a telephone network version of the identical TIMIT utterances (N-TIMIT). In this paper, we present insertion and deletion results for the segmenter (for both conditions, compared to hand transcriptions) as well as patterns observed in segmentation errors as a function of data set. Also presented will be the results of the classification algorithm for both databases.
Bibliographic reference. Chigier, Benjamin / Urdang, Erik / Spitz, Judith (1989): "Analysis of two algorithms for telephone speech recognition", In EUROSPEECH-1989, 1405-1407.