next up previous contents
Next: Task 5.2: Future Development Up: Experiments Previous: Experiments

Hidden Neural Network (HNN)

The ``hidden neural network'' (HNN) is essentially an HMM with two networks attached to each state: The network outputs are not locally normalized, so they are not formally probabilities. Each state is corresponds to a phone q.m

The model is trained by maximizing the conditional likelihood of the correct (phone) labeling q given the observation sequence x:

\begin{displaymath}P(q \vert x, M) = \frac{P(x, q \vert M)}{P(x \vert M)} \end{displaymath}

To compute the probability of the labeling requires two forward passes: one in the `clamped' phase to compute P(q | x, M), one in the 'free' phase to compute P(x | M). This quantity may be estimated using the transition and match networks attached to each state.

We have used this model for recognition experiments using the Phonebook database. the results we have obtained are slightly inferior to those obtained in SPRACH (task T4.1)using an HMM/MLP hybrid. For the 75 word task, an HNN with 37K parameters trained on 9,000 words resulted in a WER of 4.5%, compared with a WER of 1.5% for the HMM/MLP trained on 19,000 words using 166K parameters.

A relative WER reduction of 20-30% was observed with the HNN when the forward probabilities were used in decoding in place of the Viterbi criterion.

This work is reported in more detail in the ICASSP submission contained in the appendix.


next up previous contents
Next: Task 5.2: Future Development Up: Experiments Previous: Experiments
Christophe Ris
1998-11-10