This talk addresses recent activities in speech recognition and synthesis at Tokyo Institute of Technology. An important component of our speech recognition research is the investigation of Japanese broadcast news transcription and automatic topic extraction from the transcribed news speech. Since most of the Japanese words have multiple readings, we have constructed a language model that depends on the readings of words. In broadcast-news, each speaker utters several sentences in succession, therefore we applied on-line, incremental speaker adaptation after identifying the speaker of the sentence. The reading-dependent language model reduced word error rate by 4.7%, and the on-line speaker adaptation reduced word error rate by 11.8%. Topic words, which represent the content of speech, are extracted from speech recognition results based on the significance score of each word. When eight topic words are extracted for each broadcast news article (recall = 20%), 80% of them are correct on average. The topic words are used as key words to create summarizing sentences. We also propose a new formulation for speech recognition, which maximizes the a posteriori probability of the speaker's intended message for a given observed acoustic sequence. In our speech synthesis research, we are examining the use of triphone HMMs for producing cepstral parameter sequences. Experimental results indicated that natural-sounding continuous speech can be synthesized with this method.