In large-vocabulary speech recognition previous knowledge about the voice of the speaker is extremely useful for achieving high accuracy. For systems based on VQ and HMMs, a training sample is collected from the new speaker to build the VQ codebook and HMM parameters. We describe some techniques aimed at achieving a fast, adaptation of speaker-independent codebook and HMM parameters, based on a small speech sample provided by the new speaker. Results are presented concerning the 20000-word vocabulary, isolated-utterance, real-time IBM recognizer for the Italian language.