A One-Pass Real-Time Decoder Using Memory-Efficient State Network

Jian SHAO; Ta LI; Qingqing ZHANG; Qingwei ZHAO; Yonghong YAN

doi:10.1093/ietisy/e91-d.3.529

A One-Pass Real-Time Decoder Using Memory-Efficient State Network

Jian SHAO, Ta LI, Qingqing ZHANG, Qingwei ZHAO, Yonghong YAN

Full Text Views

0

Share
Cite this

Summary :

This paper presents our developed decoder which adopts the idea of statically optimizing part of the knowledge sources while handling the others dynamically. The lexicon, phonetic contexts and acoustic model are statically integrated to form a memory-efficient state network, while the language model (LM) is dynamically incorporated on the fly by means of extended tokens. The novelties of our approach for constructing the state network are (1) introducing two layers of dummy nodes to cluster the cross-word (CW) context dependent fan-in and fan-out triphones, (2) introducing a so-called "WI layer" to store the word identities and putting the nodes of this layer in the non-shared mid-part of the network, (3) optimizing the network at state level by a sufficient forward and backward node-merge process. The state network is organized as a multi-layer structure for distinct token propagation at each layer. By exploiting the characteristics of the state network, several techniques including LM look-ahead, LM cache and beam pruning are specially designed for search efficiency. Especially in beam pruning, a layer-dependent pruning method is proposed to further reduce the search space. The layer-dependent pruning takes account of the neck-like characteristics of WI layer and the reduced variety of word endings, which enables tighter beam without introducing much search errors. In addition, other techniques including LM compression, lattice-based bookkeeping and lattice garbage collection are also employed to reduce the memory requirements. Experiments are carried out on a Mandarin spontaneous speech recognition task where the decoder involves a trigram LM and CW triphone models. A comparison with HDecode of HTK toolkits shows that, within 1% performance deviation, our decoder can run 5 times faster with half of the memory footprint.

Publication: IEICE TRANSACTIONS on Information Vol.E91-D No.3 pp.529-537

Publication Date: 2008/03/01

Publicized

Online ISSN: 1745-1361

DOI: 10.1093/ietisy/e91-d.3.529

Type of Manuscript: Special Section PAPER (Special Section on Robust Speech Processing in Realistic Environments)

Category: ASR System Architecture

Cite this

Copy

Jian SHAO, Ta LI, Qingqing ZHANG, Qingwei ZHAO, Yonghong YAN, "A One-Pass Real-Time Decoder Using Memory-Efficient State Network" in IEICE TRANSACTIONS on Information, vol. E91-D, no. 3, pp. 529-537, March 2008, doi: 10.1093/ietisy/e91-d.3.529.
Abstract: This paper presents our developed decoder which adopts the idea of statically optimizing part of the knowledge sources while handling the others dynamically. The lexicon, phonetic contexts and acoustic model are statically integrated to form a memory-efficient state network, while the language model (LM) is dynamically incorporated on the fly by means of extended tokens. The novelties of our approach for constructing the state network are (1) introducing two layers of dummy nodes to cluster the cross-word (CW) context dependent fan-in and fan-out triphones, (2) introducing a so-called "WI layer" to store the word identities and putting the nodes of this layer in the non-shared mid-part of the network, (3) optimizing the network at state level by a sufficient forward and backward node-merge process. The state network is organized as a multi-layer structure for distinct token propagation at each layer. By exploiting the characteristics of the state network, several techniques including LM look-ahead, LM cache and beam pruning are specially designed for search efficiency. Especially in beam pruning, a layer-dependent pruning method is proposed to further reduce the search space. The layer-dependent pruning takes account of the neck-like characteristics of WI layer and the reduced variety of word endings, which enables tighter beam without introducing much search errors. In addition, other techniques including LM compression, lattice-based bookkeeping and lattice garbage collection are also employed to reduce the memory requirements. Experiments are carried out on a Mandarin spontaneous speech recognition task where the decoder involves a trigram LM and CW triphone models. A comparison with HDecode of HTK toolkits shows that, within 1% performance deviation, our decoder can run 5 times faster with half of the memory footprint.
URL: https://globals.ieice.org/en_transactions/information/10.1093/ietisy/e91-d.3.529/_p

Copy

@ARTICLE{e91-d_3_529,
author={Jian SHAO, Ta LI, Qingqing ZHANG, Qingwei ZHAO, Yonghong YAN, },
journal={IEICE TRANSACTIONS on Information},
title={A One-Pass Real-Time Decoder Using Memory-Efficient State Network},
year={2008},
volume={E91-D},
number={3},
pages={529-537},
abstract={This paper presents our developed decoder which adopts the idea of statically optimizing part of the knowledge sources while handling the others dynamically. The lexicon, phonetic contexts and acoustic model are statically integrated to form a memory-efficient state network, while the language model (LM) is dynamically incorporated on the fly by means of extended tokens. The novelties of our approach for constructing the state network are (1) introducing two layers of dummy nodes to cluster the cross-word (CW) context dependent fan-in and fan-out triphones, (2) introducing a so-called "WI layer" to store the word identities and putting the nodes of this layer in the non-shared mid-part of the network, (3) optimizing the network at state level by a sufficient forward and backward node-merge process. The state network is organized as a multi-layer structure for distinct token propagation at each layer. By exploiting the characteristics of the state network, several techniques including LM look-ahead, LM cache and beam pruning are specially designed for search efficiency. Especially in beam pruning, a layer-dependent pruning method is proposed to further reduce the search space. The layer-dependent pruning takes account of the neck-like characteristics of WI layer and the reduced variety of word endings, which enables tighter beam without introducing much search errors. In addition, other techniques including LM compression, lattice-based bookkeeping and lattice garbage collection are also employed to reduce the memory requirements. Experiments are carried out on a Mandarin spontaneous speech recognition task where the decoder involves a trigram LM and CW triphone models. A comparison with HDecode of HTK toolkits shows that, within 1% performance deviation, our decoder can run 5 times faster with half of the memory footprint.},
keywords={},
doi={10.1093/ietisy/e91-d.3.529},
ISSN={1745-1361},
month={March},}

Copy

TY - JOUR
TI - A One-Pass Real-Time Decoder Using Memory-Efficient State Network
T2 - IEICE TRANSACTIONS on Information
SP - 529
EP - 537
AU - Jian SHAO
AU - Ta LI
AU - Qingqing ZHANG
AU - Qingwei ZHAO
AU - Yonghong YAN
PY - 2008
DO - 10.1093/ietisy/e91-d.3.529
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E91-D
IS - 3
JA - IEICE TRANSACTIONS on Information
Y1 - March 2008
AB - This paper presents our developed decoder which adopts the idea of statically optimizing part of the knowledge sources while handling the others dynamically. The lexicon, phonetic contexts and acoustic model are statically integrated to form a memory-efficient state network, while the language model (LM) is dynamically incorporated on the fly by means of extended tokens. The novelties of our approach for constructing the state network are (1) introducing two layers of dummy nodes to cluster the cross-word (CW) context dependent fan-in and fan-out triphones, (2) introducing a so-called "WI layer" to store the word identities and putting the nodes of this layer in the non-shared mid-part of the network, (3) optimizing the network at state level by a sufficient forward and backward node-merge process. The state network is organized as a multi-layer structure for distinct token propagation at each layer. By exploiting the characteristics of the state network, several techniques including LM look-ahead, LM cache and beam pruning are specially designed for search efficiency. Especially in beam pruning, a layer-dependent pruning method is proposed to further reduce the search space. The layer-dependent pruning takes account of the neck-like characteristics of WI layer and the reduced variety of word endings, which enables tighter beam without introducing much search errors. In addition, other techniques including LM compression, lattice-based bookkeeping and lattice garbage collection are also employed to reduce the memory requirements. Experiments are carried out on a Mandarin spontaneous speech recognition task where the decoder involves a trigram LM and CW triphone models. A comparison with HDecode of HTK toolkits shows that, within 1% performance deviation, our decoder can run 5 times faster with half of the memory footprint.
ER -