@mastersthesis{
    smit2011stackedtransformsthesis,
    author = "Smit, Peter",
    school = "Aalto University School of Science",
    language = "eng",
    title = "Stacked transformations for foreign accented speech recognition",
    url = "http://users.ics.tkk.fi/peter/papers/thesis.pdf",
    abstract = "Nowadays, large vocabulary speech recognizers exist that are performing reasonably well for specific conditions and environments. When the conditions change however, performance degrades quickly. For example, when the person to be recognized has a foreign accent the conditions could mismatch with the model, resulting in high error rates. The problem in recognizing foreign accented speech is the lack of sufficient training data. If enough data would be available of the same accent, from numerous different speakers, a well performing accented speech model could be built. Besides the lack of speech data, there are more problems with training a complete new model. It costs a lot of computational resources and storage space to train a new model. If speakers with different accents must be recognized, these costs explode as every accent needs retraining. A common solution for preventing retraining is to adapt (transform) an existing model, such that it better matches the recognition conditions. In this thesis multiple different adaptation transformations are considered. Speaker Transformations are using speech data from the target speaker, Accent Transformations use speech data from different speakers, who have the same accent as the speech that needs to be recognized. Neighbour Transformations are estimated with speech from different speakers that are automatically determined to be similar to the target speaker. Novelty in this work is the stack wise combination of these adaptations. Instead of using a single transformation, multiple transformations are `stacked together'. Because all adaptations except the speaker specific adaptation can be precomputed, no extra computational costs at recognition time occur compared to normal speaker adaptation and the adaptations that can be precomputed are much more refined as they can use more and better adaptation data. In addition, they need only a very small amount storage space, compared to a retrained model. The effect of Stacked Transformations is that the models have a better fit for the recognition utterances. When compared to no adaptation, improvements up to 30\% in Word Error Rate can be achieved. In adaptation with a small number (5) of sentences, improvements up to 15\% are gained.",
    pdf = "thesis.pdf",
    month = "May",
    responsibleauthor = "Smit, Peter",
    flags = "AIRC",
    il = "no",
    year = "2011",
    keywords = "automatic speech recognition, foreign accent recognition, linear transformation, stacked transformation",
    unitcode = "T306",
    impactfactor = "G2"
}