Architecture of an automatic speech recognition system
Today, state-of-the-art systems for automatic speech recognition are
based on the statistical approach of Bayes decision rule. The
implementation of Bayes decision rule for automatic speech recognition
is based on two kinds of stochastic models: the acoustic model and the
language model which together are the basis for the decision process
itself, i.e. the search for the most probable sentence. These modules
of an automatic speech recognition system
are characterized as follows:
The acoustic model captures the acoustic properties of
speech and provides the probability of the observed acoustic
signal given a hypothesized word sequence. The acoustic model
includes:
The acoustic analysis which parameterizes the speech
input into a sequence of acoustic vectors.
Acoustic models for the smallest sub-word units,
i.e. phonemes which usually are modelled context
dependent.
The pronunciation lexicon, which defines the
decomposition of the words into the subword units.
Topology and search space for a
Hidden Markov Model (HMM) for the word
"sieben"
Speech waveform of the utterance
"Sollen wir am Sonntag nach Berlin fahren", and the corresponding
FFT spectrum
The language model captures the linguistic
properties of the language and provides the a-priori probability
of a word sequence. From an information theoretic point of view,
syntax, semantics, and pragmatics of the language could also be
viewed as redundancies. Because of the stochastic nature of such
redundancies, language models usually are based on
statistical concepts.
Search realizes Bayes decision criterion on the basis
of the acoustic model and the language model. This requires the
generation and scoring of competing sentence hypotheses. To
obtain the final recognition result, the main objective then is
to search for that sentence hypothesis with the best score using
dynamic programming. The efficiency of the search process is
increased by pruning unlikely hypotheses as early as possible
during dynamic programming without affecting the recognition
performance.
Last modified November 2, 2001
