This paper deals with adaptation to new test environments of an HMM recognizer trained in a given condition. In particular, a speaker-independent continuous speech recognizer has been trained using a multispeaker database collected through a telephone with an electret transducer and tested with several types of hand-sets having different spectral response characteristics. The results of these experiments show that system performance drops dramatically on test utterances collected through carbon transducers. A procedure for hierarchical spectral mapping, originally proposed for speaker adaptation, is used f0 channel adaptation. A codebook adapted to the new channel is obtained from a database including a limited number of adaptation sentences. Test sentences are then vector quantized using the adapted codebook. Tested with sentences acquired through a carbon transducer, more than 60% of performance improvement is achieved, while nearly 100% of the errors can be recovered when electret type transducers are used for testing.

Adapted HMMs are then estimated by aligning the sentences collected in the new environment against the HMMs trained in the original environment, and nearly 92% of the errors are recovered in the carbon transducer test.