CMLLR and semicontinuous models

[Anonymous]

Is CMLLR adaptation (using SphinxTrain/python/cmusphinx/cmllr.py) supposed
to work with semicontinuous and PTM models? Currently, with both types, I get
an error:

raceback (most recent call last):
  File "/home/tanel/devel/cmusphinx/SphinxTrain/python/cmusphinx/cmllr.py", line 305, in <module>
    Ws = estimate_cmllr(stats, inmean, invar, mdef)
  File "/home/tanel/devel/cmusphinx/SphinxTrain/python/cmusphinx/cmllr.py", line 97, in estimate_cmllr
    Ginv = np.linalg.inv(G[i])
  File "/usr/lib/pymodules/python2.6/numpy/linalg/linalg.py", line 423, in inv
    return wrap(solve(a, identity(a.shape[0], dtype=a.dtype)))
  File "/usr/lib/pymodules/python2.6/numpy/linalg/linalg.py", line 306, in solve
    raise LinAlgError, 'Singular matrix'
numpy.linalg.linalg.LinAlgError: Singular matrix

CMLLR with continuous models works perfectly.

[Nickolay V. Shmyrev]

Hello

The MLLR adaptation has little sense for semi-continuous models where the most
information is stored in mixture weights.

Also, the cmllr.py doesn't support multistream features, though one can add
this support.

[Anonymous]

Thanks, it didn't occur to me that CMLLR doesn't make much sense for semi-
continuous models.

I wonder what should I then use for unsupervised adaptation of sem-continuous
models. CMLLR and MLLR do not make sense, and this
document specifically
says not to use MAP for unsupervised adaptation.

[Nickolay V. Shmyrev]

and this document specifically says not to use MAP for unsupervised
adaptation.

It's not quite accurate statement. It all depends on the amount of data. If
you have large data and use small smoothing factor (large tau) the adapted
model should be good enough.

It's all about using suffciient data to estimate the parameters. There are
specific methods to deal with small amount of data, for example structured MAP
(SMAP) adaptation, but I'm not sure how does it apply to semi-continuous case.
See also

CROSSLINGUAL ADAPTATION OF SEMI-CONTINUOUS HMMS USING ACOUSTIC SUB-SIMPLEX
PROJECTION
Frank Diehl, Asuncion Moreno, Enric Monte

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.74.7313&rep=rep1&typ
e=pdf