Long Audio Alignment: Dynamic Linguist and Phone Loops

My last update on Phone Loops indicated some significant improvements in Out of Grammar words recognition while aligning audio files with text, but it also remarked on alignment of large audio file's  need for some special attention because of the huge search space generated the linguist and it's associated memory.

I hence modified the linguist to dynamically generate search graph during the process of recognition by adding Phone Loop only at the current grammar state, hence significantly reducing the memory required to store 1 phone-loop per word in the transcription. Some tests were conducted as well to determine 'Out of Grammar Branch" probability and 'Phone Insertion Probability" for audio files with some noise and close to 3% error in transcription. Best results in Out of Grammar word recognition were achieved at out of grammar branch probability ~ 1E-8 and phone insertion probability ~ 1E - 80. The word error rate hence obtained was close 5%.

Current version of the linguist is now available in long-audio-aligner branch. We now proceed towards the implementation of Keyword spotting algorithm for improving the alignment even further.

Long Audio Training - reduced B-W computation & move towards CUDA

It's been a while since my last post. In theese days I was modifying the Baum-Welch algorithm to the reduced version, which is finally complete.

Forward, backward and Viterbi methods were changed in the following way:

  • 'Reduced' forward method was created. This method computes the checkpoints for later re-computation of actual alpha & scaling values. The size of reduced matrices is a function of block size, which is taken as a parameter.
  • 'Local' forward method was created. This method performs the alpha values re-computation for a particular checkpoint (block of values).
  • As SphinxTrain has Viterbi back-pointers computation embedded in the Forward pass, the modification of Viterbi was just to use the reduced forward and to recompute the alpha values with the local forward.
  • Backward update was modified in a similar way as the Viterbi.

The modification was successfully tested on an4 database. It performed somewhat slower, which was anticipated, as the modified algorithm does more computation.

I also tried the modification on the 'rita' (long audio) database. I was forced to quit the computation as it took all my system's memory. This sadly seems as no improvement in the memory demands and might suggest that some of the memory demands are not in the forward/backward/Viterbi as well as that I might have just introduced some memory leaks. During the brief tests the block_size parameter was set arbitrarily to 11, not the sqrt function of time frames count, which also may have some performance consequences.

The actual slow-down and memory requirements are subject to more detailed tests.

Regarding the CUDA, I have gain access to 3 CUDA machines. Two of them belong to Sitola, The Laboratory of Advanced Network Technologies. The access to these machines is provided by MetaCentrum, Czech academic grid organization providing the computation and storage resources. The cards are GeForce GTX 285, GeForce 8800 Ultra and GeForce 8400M GS (a rather low-end one in my personal laptop). These are devices the CUDA development and testing will take place on. More info to come, please check the project page.

Long Audio Alignment: Phone Loops

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) .

Long Audio Alignment: Week 4

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.