LDA/MLLT problem

[mchammer]

Hi,

training acoustic models with parameter $CFG_LDA_MLLT = 'yes' gives varying results using the script RunAll.pl (the same for RunAll_CDMLLT.pl) on every clean run (generated model and inter-model files are deleted after each run) I perform. I'm using always the same feature files I created once and am not changing any parameters.
Training without LDA/MLLT switch on, gives always get the same results.
I tried this on an4 and rm1 corpora as described in the tutorial http://www.speech.cs.cmu.edu/sphinx/tutorial.html.
I'm using sphinxbase-0.4.1, SphinxTrain-1.0 and sphinx3-0.8.

Some results for the an4 corpus, each line corresponds to a different run:
SENTENCE ERROR: 53.1% (69/130) WORD ERROR RATE: 16.4% (126/773)
SENTENCE ERROR: 50.8% (66/130) WORD ERROR RATE: 16.3% (125/773)
SENTENCE ERROR: 47.7% (62/130) WORD ERROR RATE: 14.6% (112/773)
SENTENCE ERROR: 55.4% (72/130) WORD ERROR RATE: 17.5% (135/773)
SENTENCE ERROR: 51.5% (66/130) WORD ERROR RATE: 16.6% (128/773)
SENTENCE ERROR: 50.0% (65/130) WORD ERROR RATE: 15.5% (119/773)
SENTENCE ERROR: 56.2% (73/130) WORD ERROR RATE: 16.9% (130/773)
SENTENCE ERROR: 62.3% (80/130) WORD ERROR RATE: 22.3% (172/773)
SENTENCE ERROR: 56.2% (73/130) WORD ERROR RATE: 17.3% (133/773)
SENTENCE ERROR: 54.6% (70/130) WORD ERROR RATE: 16.2% (125/773)
SENTENCE ERROR: 53.1% (69/130) WORD ERROR RATE: 17.3% (133/773)
SENTENCE ERROR: 60.8% (79/130) WORD ERROR RATE: 21.9% (169/773)

Is this a bug, or is there an explanation for this behavior.

Regards,

Chris

[Nickolay V. Shmyrev]

Ah wait, mllt.py uses random:

def train(self, A=None):
    """ 
    Train an MLLT transform from an optional starting point.

    @param A: Initial MLLT matrix to start training. 
    @type A: numpy.ndarray 
    @return: Optimized MLLT transformation matrix 
    @rtype: numpy.ndarray 
    """
    if A == None:
        # Initialize it with a random positive-definite matrix of 
        # the same shape as the covariances 
        s = self.cov[0].shape
        d = -1
        while d < 0:
            A = eye(s[0]) + 0.1 * random(s)
            d = det(A)

so it could be explained. But I'm not happy with such a huge accuracy jumps anyhow.

[Nickolay V. Shmyrev]

Hm, indeed so. It seems that mllt computation is unstable. The first thing that is different during training attempts is a mllt matrix and the log of mllt computation in 02.mllt_train. Looks like the serious issue.