ADAPTIVE TECHNIQUES - MINIMIZING MISMATCH
- We can improve recognition performance by training on a single
speaker. This is known as speaker dependent speech recognition.
- However, there are numerous training problems (long enrollment).
An alternate approach is to adapt speaker independent models.
- Such adaptation techniques are generally used to reduce mismatch
between the acoustic models and the decoding environment (e.g.,
microphone, acoustic channel and speaker mismatch).
- There are two basic approaches:
- Maximum A Posteriori (MAP): choosing an estimate
that maximizes the posterior probability (consistent
with the observed data and prior inormation).
- Maximum Likelihood Linear Regression (MLLR): ML
estimate of a linear transformation.