Old Speech Group Projects

Previous Speech Group projects include the following:

• Auditory Basis of Speech Perception - Studies of the temporal and spectral characteristics of human speech perception.
• PHiPAC (Portable High Performance ANSI C) - a set of programs for generating high performance matrix/vector functions in C.
• REMAP (Recursive Estimation and Maximization of A posteriori Probabilities) - an iterative procedure for training and using MLPs for speech recognition and confidence estimation based on a posteriori probabilities.
• Syllable-based ASR - a project to improve speech recognition by incorporating syllable-length acoustic representations and models.
• THISL (THematic Indexing of Spoken Language) - a European ESPRIT project (with ICSI as a subcontractor) researching the retrieval of multimedia information from errorful text that has been transcribed from broadcast programs using ASR.
• BMM (Buried Markov Models) - extending HMMs to include explicit dependencies between observation vectors.
• DPM (Dynamic Pronunciation Modeling) - modeling pronunciation incorporating contextual factors such as speaking rate and word sequence probability.
• LDLM (Long Distance Language Models) - extending LMs to include dependencies on hidden topic sources.
• LVCSR (Large Vocabulary Continuous Speech Recognition) - applying robust features, hybrid HMM/connectionist techniques, and pronunciation modeling to large vocabulary continuous speech recognition, particularly for broadcast news and spontaneous conversations.
• MSG (Modulation-filtered SpectroGram) - a perceptually-inspired robust front end for speech recognition systems.
• LVCSR (Large Vocabulary Continuous Speech Recognition) - applying robust features, hybrid HMM/connectionist techniques, and pronunciation modeling to large vocabulary continuous speech recognition, particularly for broadcast news and spontaneous conversations.