Presentation | 2018-07-26 Knowledge Distillation from Neural Network Based Acoustic Model based on Different Decision Tree Takashi Fukuda, Samuel Thomas, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | This paper proposes a method to transfer acoustic knowledge from teacher network with a different decision tree to a student network. The teacher model has different output layer from the student network but has high recognition performance. In the proposed method, (1) phone alignments are generated from each phoneme context dependent decision tree, which relates to the output layer of the network, and (2) create a confusion matrix representing a relationship of phoneme contexts between teacher's and student's output nodes. In the experiments, we show that the proposed method contributes to the improvement of student network and report that 9.6% relative improvement was obtained by the proposed method over the acoustic model constructed only with hard labels on noisy environment speech recognition task with Aurora 4. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Speech recognition / acoustic model / knowledge distillation / decision tree / phone mapping |
Paper # | SP2018-20 |
Date of Issue | 2018-07-19 (SP) |
Conference Information | |
Committee | SP / IPSJ-SLP |
---|---|
Conference Date | 2018/7/26(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Sago-Royal-Hotel (Hamamatsu) |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Speech recognition and understanding, dialog system, etc. |
Chair | Yoichi Yamashita(Ritsumeikan Univ.) / Masafumi Nishimura(Shizuoka Univ.) |
Vice Chair | Akinobu Ri(Nagoya Inst. of Tech.) |
Secretary | Akinobu Ri(Kyoto Univ.) / (Meijo Univ.) |
Assistant | Tomoki Koriyama(Tokyo Inst. of Tech.) / Satoshi Kobashikawa(NTT) |
Paper Information | |
Registration To | Technical Committee on Speech / Special Interest Group on Spoken Language Processing |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Knowledge Distillation from Neural Network Based Acoustic Model based on Different Decision Tree |
Sub Title (in English) | |
Keyword(1) | Speech recognition |
Keyword(2) | acoustic model |
Keyword(3) | knowledge distillation |
Keyword(4) | decision tree |
Keyword(5) | phone mapping |
1st Author's Name | Takashi Fukuda |
1st Author's Affiliation | IBM Japan(IBM) |
2nd Author's Name | Samuel Thomas |
2nd Author's Affiliation | IBM T. J. Watson Research Center(IBM) |
Date | 2018-07-26 |
Paper # | SP2018-20 |
Volume (vol) | vol.118 |
Number (no) | SP-160 |
Page | pp.pp.21-24(SP), |
#Pages | 4 |
Date of Issue | 2018-07-19 (SP) |