Time-first search for large vocabulary speech recognition

A fast HMM match algorithm for very large vocabulary speech recognition

“…4. The method can be integrated with most existing HMM-based speech decoding principles including DP searches (Ney and Ortmanns, 1999), time-synchronous algorithms (Odell, 1995), stack decoders (Jelinek, 1969;Paul, 1992;Robinson and Christie, 1998;Renals and Hochberg, 1999), with heuristic methods; e.g. A Ã /IDA Ã (Hart et al, 1968;Kenny et al, 1993;Korf, 1985) and finite-state-transducer methods (Aubert, 2000;Seward, 2001;Mohri et al, 2002).…”

An overview of decoding techniques for large vocabulary continuous speech recognition

“…This fast-match component typically relies on a lexical prefix tree and on simplified acoustic models to achieve the highest efficiency. 8 Concerning the start-synchronous lexical tree exploration, it is worth pointing out that this step has been achieved either with a standard time-synchronous DP scheme (Hochberg et al, 1994) or with a "time-first" asynchronous method (Robinson & Christie, 1998), the latter requiring less memory storage.…”

The THISL system for indexing and retrieval of broadcast news

“…For this work, the CHRONOS decoder [5] has been developed to allow memory efficient decoding of whole broadcasts.…”

“…The CHRONOS decoder [5] has been developed to allow the decoding of an entire show in real-time. Using a 450MHz Pentium-I1 running Solaris, it was possible to achieve real-time decoding with a typical memory usage of under 256Mb.…”