Our objective this week was to model presence of words in the utterance that are not in the transcription. The approach used was to model it using Phone Loops. A phone loop contains all the phones of an acoustic model and can model any utterance (i.e. also words in the transcription).  Hence the key to good alignment using phone loop is an optimal branch probability which is large enough so that recogniser does not mistake a OOV word as a word in the grammar and small enough to not replace a word in grammar by a OOV word.

A linguist satisfying the above criteria has been added to long audio aligner branch. However, the linguist performs quite well for small sized transcriptions , the size of the search graph produced is too large for small sized transcription. We plan to generate this search graph dynamically now, to solve this memory issue. This way the memory requirements for generating and storing huge search graphs will be reduced to almost  O(1) .

Extending my work from last week, I have added a framework for testing audio aligner. The framework includes tools to read transcription from a file and convert it into a format compatible with grammar, corrupt a transcription with desired error rates of particular types and finally a pronunciation generator for OOV words.

The problem we are trying to tackle first is of word insertions. For this, the aligner grammar has been modified to contain word self-loops and one backward jump per word. The word error rates for small audio (~ 5 min) with word repetitions is less than 2% with the current configuration. However it models word repetitions quite well, it can't model word insertion completely as words inserted don't necessarily come from the neighborhood of recently decoded word.
A similar (but not quite the same) out of grammar utterance recognition problem was attempted in S4 by adding an out of grammar branch (PhoneLoops) in SentenceHMM parallel to the one obtained by expanding grammar. But this addition does not model insertion of a couple of words in an utterance that "almost" fits the grammar.
So I plan to modify the same approach, but rather than having a branch parallel to one generated from grammar, I will have similar branches between consecutive words nodes in the grammar. The branch out probability for this branch will require some experimentation as well, but I expect this addition to model insertion of words completely.

During the last week the memory management of the current SphinxTrain version was examined using Valgrind family tools.

Program was profiled on one pass Baum-Welch training on the whole an4 database (948 files) and on one recording from custom 'rita' database. The total length of traing recordings was ~42 minutes and 5:31 minutes respectively. The reason for the testing on rita database considered only one file was the big amount of time taken by the training process. In both cases this was about 8 minutes without Valgrind profiling. (This time has increased about 15 times when profiling.)

First interesting result is that the training on 42 minutes of short audio files (each few seconds long) took about the same as training on a single file 5:31 minutes long.

During the experiments an issue with extensive memory reallocation was found in the forward algorithm. This was cca 760,000 of __ckd_realloc__ function calls when training on an4 database. The cause was found by the analysis of source codes: initial active state set allocation size was repeatedly set too low and after that vastly insreased (by about 1.3 GB on rita database) by the step of a few bytes. Algorithm was modified to reallocate the memory by quadratic step instead of linear. The modification reduced the reallocation to the 18,500 function calls on an4 database, thus by the factor of about 40. The tradeof was slight increase in the total amount of memory allocated from 5.7 MB to 6.2 MB and from 1.4 GB to 1.7 GB on an4 and rita databases respectively. The multiplication factor was set to 2 but this was arbitrary and can be subject to optimization.

In theory this modification could also reduce the running time of the algorithm but significant reduce of time was not measured (approximately 8 minutes +- few seconds in all cases - an4 & rita with and without the modification).

A question was whether a substantial fraction of memory demands could be attributed to memory leaks. This has shown not to be the case. Some memory leaks were found but nothing significant - under 1 MB in all cases. It was concluded that the reduced space technique implementation is necessary.

The work on the implementation of reduced space Baum-Welch has begun. This will involve a major rewriting of the forward, backward and Viterbi algorithms using checkpoints method (as described in article http://www.cse.ucsc.edu/research/compbio/papers/samspace.pdf). This method allows parametrization reducing the memory use from linear with respect to T to an arbitrary integer-root of T. The tradeof is increase in the computation length by the same factor. During this also the memory leaks should be completely prevented.

In order of testing this should be done in few phases to ensure the new implementation is correct. More information will be found at the project wiki https://cmusphinx.github.io/wiki/longaudiotraining as the work progresses.

Last update indicated an error rate of almost 16 % for sufficiently large audio files. Experiments were conducted to pin-point the source of these errors.  It was then suggested to align audio without classifying speech and non-speech components of the audio.  Alignment with such configuration is now being tested with different grammars for different sorts of possible errors in the audio and/or transcription.

Audio and it's perfect transcription for up to 20 minutes long utterances have been checked for alignment with the current state of aligner, and the resulted in close 0% word error rate. The grammar used for this sort of  alignment only allows transitions from word to it's immediate successor in the transcription.

We are currently classifying different sorts of errors in transcription and utterance , and modelling grammar to allow alignment for the same.