Next: Integrating Artificial Neural Networks
Up: Speech Recognition with HMMs
Previous: Imbalance Between Emission and
In this section, we briefly discuss the main unrealistic probabilistic
assumptions behind state-of-the-art speech recognition systems based on HMMs.
Although these problems have been known for many years, and much research has
been done to address these problems, this research has not yielded
significantly improved recognition performance.
- The HMM topology (which is very sparse) imposes a very
strong a-priori on the meaning of the state variable, which is
typically not realistic. In particular it does not take into
account many possible effects of context (and this is related
to several of the problems below).
- Some important long-term structure is not taken into account,
such as, the identity of the speaker (which induces a dependency
on the consecutive observations from the same speaker), and prosody
(which represents the slow-varying sequential structure of pitch,
duration, and energy).
- The modeling of durations is generally not realistic
(see also the problem of imbalance between emission and
transition probabilities which results in speech recognition
not taking much account of the duration model incorporated
in non-zero transition probabilities, section 3.6).
- Consecutive acoustic vectors are not conditionally independent
(given the state). This is in fact often forced on the
data sequence
by using in the observation vector for each time step information
that overlaps with that in neighboring time steps (e.g., using first
and second time-derivatives of the spectral or cepstral coefficients).
- The acoustic features used as observations do not contain
all the relevant information from the speech (typically most systems
use cepstral coefficients, obtained from a short-term power spectrum).
Yoshua Bengio
Tue Oct 7 08:34:36 EDT 1997