摘要 |
PURPOSE: A method for adapting compressed acoustic model is provided to adapt acoustic models that are quantized and compressed using adaptation training data in multi-dimensional sub-spaces. CONSTITUTION: Acoustic models quantized in sub-spaces are adapted using adaptation training data. The acoustic models are compressed into multi-dimensional sub-spaces. A codebook is generated for each of the sub-spaces. Adaptation transform is estimated and applied to codewords of codebooks. The codebook is divided into regression classes. The regression classes reflect whether Gaussian distribution belongs to a language class such as a consonant, a vowel or mute.
|