Pedro J. Moreno, "Speech Recognition in Telephone Environments", Master's Report, ECE Department, CMU, January, 1993. Abstract This work compares the automatic recognition of telephone-quality speech with that of speech recorded in noiseless and acoustically controlled environments. The report includes a description of the SPHINX speech recognition system, a general overview of the field of robust speech recognition, transmission characteristics of the telephone network, a description of the speech databases used, and the speech recognition experiments that have been done. Speech recognition systems work very well dealing with high-quality speech; however, when the speech signal has been corrupted by the telephone network, recognition accuracies decrease dramatically. In the databases we use in this research, a relative increase of about 20% in recognition error rate occurred when telephone speech was used. We study some possible factors that could account for this loss. Bandwidth reduction and the presence of low-frequency tones can explain part of the loss but not all of it. A potentially improper tuning of front-end signal processing parameters is not the problem either. Preliminary results show that the standard model of environmental degradation that has successfully improved the recognition accuracy of SPHINX in office environments does not sufficiently describe the degradation that is introduced by transmission over a telephone channel. Environmental normalization algorithms such as CDCN (Codeword Dependent Cepstral Normalization) and RASTA (Relative Spectral Processing) that assume this simple model do not provide complete compensation.