Integration of acoustic information and PPRLM scores in a multiple-Gaussian classifier for Language Identification

Abstract

In this paper, we present several innovative techniques that can be applied in a PPRLM system for language identification (LID). We will show how we obtained a 535%0 relative error reduction from our base system using several techniques. First, the application of a variable threshold in score computation, dependent on the average scores in the language model, provided a 35% error reduction. A random selection of sentences for the different sets and the use of silence models also improved the system. Then, to improve the classifier, we compared the bias removal technique (up to 19% error reduction) and a Gaussian classifier (up to 37%0 error reduction). Finally, we included the acoustic score in the Gaussian classifier (2% error reduction) and increased the number of Gaussians to have a multipleGaussian classifier (14% error reduction). We will show how all these improvements are remarkable as they have been mostly additive.

Language identification based on n-gram frequency ranking

“…The language with fewer units will have higher probabilities in the LM score, and so the classifier will tend to select that language. To tackle this issue, we proposed in [3] to use a Gaussian classifier instead of the usual decision formula applied in PPRLM. With all the scores provided by every LM in the PPRLM module we prepare a score vector.…”

“…To estimate the Gaussian distribution we used the acoustic models training list, as this data does not participate in the LM estimation. We demonstrated in [3] that it was a good option in order to make a better use of the training list.…”

“…To increase the number of Gaussians we have followed the classical HMM modeling approaches (Gaussian splitting and Lloyd reestimation after each splitting). More details of our system can be found in [2]and [3]. In Table 2 we can see the results using PPRLM and our Multiple-Gaussian classifier for the optimum combination of weights.…”

“…The most used technique in Language identification (LID) is the phone-based approach, like Parallel phone recognition followed by language modeling (PPRLM) [1]- [3], which classifies languages based on the statistical characteristics of the allophone sequences with a very good performance.…”

“…PPRLM does not model long-span dependencies. As we checked during the work carried out in [2] and [3], best results can be obtained using trigrams, but with 4-grams language models results are slightly worse, probably due to insufficient training data to estimate them reliably.…”