US5349645A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Word hypothesizer for continuous speech decoding using stressed-vowel centered bidirectional tree searches
Assignee: Panasonic Holdings Corp
•
Filing: 1991-12-31
•
Granted: 1994-09-20
•
Expires: 2011-12-31
Abstract
Abstract A word hypothesis module for speech decoding consists of four submodules: vowel center detection, bidirectional tree searches around each vowel center, forward-backward pruning, and additional short words hypotheses. By detecting the strong energy vowel centers, a vowel-centered lexicon tree can be placed at each vowel center and searches can be performed in both the left and right directions, where only simple phone models are used for fast acoustic match. A stage-wise forward-backward technique computes the word-beginning and word-ending likelihood scores over the generated half-word lattice for further pruning of the lattice. To avoid potential miss of short words with weak energy vowel centers, a lexicon tree is compiled for these words and tree searches are performed between each pair of adjacent vowel centers. The integration of the word hypothesizer with a top-down Viterbi beam search in continuous speech decoding provides two-pass decoding which significantly reduces computation time.
Inventors
- 👤 Yunxin Zhao
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-10-30 22:22
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/08; G10L15/14; G10L15/142
Claims Analysis
Total Claims: 33
Claim 1
Independent Claim
76 words
A high-speed continuous speech-decoding system decoding a speech sentence by utilizing two-pass decoding, said system comprising: means for converting a speech utterance into a sequence of feature vectors; means for detecting stressed vowel centers in said sequence of feature vectors; means for generating a word lattice based on the detected stressed vowel centers; and means for performing a time-synchronous Viterbi beam search using the word lattice from said generating means to constrain said Viterbi beam search.
Claim 2
Dependent Claim
76 words
A high-speed continuous speech-decoding system decoding a speech sentence by utilizing two-pass decoding, said system comprising: means for converting a speech utterance into a sequence of feature vectors; means for detecting stressed vowel centers in said sequence of feature vectors; means for generating a word lattice based on the detected stressed vowel centers; and means for performing a time-synchronous Viterbi beam search using the word lattice from said generating means to constrain said Viterbi beam search.
Claim 3
Dependent Claim
51 words
The continuous speech-decoding system of claim 1 wherein said means for generating a word lattice comprises: means for performing a time-synchronous bidirectional tree search around each detected stressed vowel center producing word hypotheses; and means for pruning the word hypotheses resulting from said bidirectional tree search in forward and backward directions.
Claim 4
Dependent Claim
26 words
The continuous speech-decoding system of claim 2 wherein said means for generating a word lattice further comprises: means for performing a tree search for short words.
Claim 5
Dependent Claim
30 words
The continuous speech-decoding system of claim 3 wherein said means for performing a short word tree search, searches between pairs of adjacent vowel centers using a short word dictionary tree.
Claim 6
Dependent Claim
66 words
The continuous speech-decoding system of claim 2 wherein said bidirectional tree search means comprises means for calculating the likelihood scores of a match between the feature vectors around the stressed vowel centers and a predetermined set of phone models; and said bidirectional tree search also comprises means for modeling each stressed vowel by a mixture density of a vowel phone model and a liquid-vowel syllable model.
Claim 7
Dependent Claim
19 words
The continuous speech-decoding system of claim 5 further comprising means utilizing two threshold conditions to discard low likelihood scores.
Claim 8
Dependent Claim
35 words
The continuous speech-decoding system of claim 6 wherein the first threshold condition of said two threshold discarding means is based on a maximum likelihood value from a current tree search at the current time frame.
Claim 9
Dependent Claim
34 words
The continuous speech-decoding system of claim 7 wherein the second threshold condition of said two threshold discarding means is based on the accumulated maximum likelihood score without tree path constraint of each time frame.
Claim 10
Dependent Claim
23 words
The continuous speech-decoding system of claim 8 wherein a new node word hypothesis is accepted if said first and second conditions are satisfied.
Claim 11
Dependent Claim
31 words
The continuous speech-decoding system of claim 9 wherein said new node word hypothesis comprises a node occurrence time and associated likelihood scores for either a left or a right half-word hypothesis.
Claim 12
Dependent Claim
34 words
The continuous speech-decoding system of claim 10 comprising means for comparing the node word hypothesis of each time frame from all vowel centers and keeping only the node word hypotheses with high likelihood scores.
Claim 13
Dependent Claim
56 words
The continuous speech-decoding system of claim 11 further comprising a forward-backward pruning means which operates from the first vowel center in said sequence of feature vectors to the last vowel center in said sequence of feature vectors for the forward direction and from the last vowel center to the first vowel center for the backward direction.
Claim 14
Dependent Claim
87 words
The continuous speech-decoding system of claim 12 wherein said forward-backward pruning means operates in the forward direction according to the steps: (a) calculate the maximum likelihood scores of all possible paths that end with a stressed vowel for a certain word hypothesis; (b) calculate the maximum likelihood scores of all paths that end with said certain word hypothesis; and (c) sort all the maximum likelihood scores from step (b) into descending order and keep only the paths having the maximum likelihood scores that exceed a predetermined threshold.
Claim 15
Dependent Claim
19 words
The continuous speech-decoding system of claim 1 wherein said time-synchronous Viterbi beam search comprises a top-down Viterbi beam search.
Claim 16
Dependent Claim
66 words
The continuous speech-decoding system of claim 1 wherein said vowel center detecting means comprises means for selecting only those energy peaks represented by the feature vectors which are above a predetermined energy threshold or the feature vectors associated with the energy peaks that are acoustically matched to vowels and passing indices of the time locations of the selected energy peaks to said bidirectional tree search means.
Claim 17
Dependent Claim
59 words
In a high-speed continuous speech-decoding system for decoding speech sentences into decoded word strings, said system including means for converting a speech utterance into a sequence of feature vectors, the improvement therein comprising: means for detecting stressed vowel centers in said sequence of feature vectors; and means for generating a word lattice based on the detected stressed vowel centers.
Claim 18
Dependent Claim
59 words
In a high-speed continuous speech-decoding system for decoding speech sentences into decoded word strings, said system including means for converting a speech utterance into a sequence of feature vectors, the improvement therein comprising: means for detecting stressed vowel centers in said sequence of feature vectors; and means for generating a word lattice based on the detected stressed vowel centers.
Claim 19
Dependent Claim
49 words
The improvement of claim 16 wherein said means for generating a word lattice comprises: means for performing a time-synchronous bidirectional tree search around each detected stressed vowel center producing word hypotheses; and means for pruning the word hypotheses resulting from said bidirectional tree search in forward and backward directions.
Claim 20
Dependent Claim
24 words
The improvement of claim 17 wherein said means for generating a word lattice further comprises: means for performing a tree search for short words.
Claim 21
Dependent Claim
27 words
The improvement of claim 18 wherein said short word tree search means searches only between each pairs of adjacent vowel centers using a short word dictionary tree.
Claim 22
Dependent Claim
64 words
The improvement of claim 17 wherein said bidirectional tree search means comprises: means for calculating the likelihood scores of a match between the feature vectors around the stressed vowel centers and a predetermined set of phone models; and said bidirectional tree search also comprises means for modeling each stressed vowel by a mixture density of a vowel phone model and a liquid-vowel syllable model.
Claim 23
Dependent Claim
16 words
The improvement of claim 20 further comprising means utilizing threshold conditions to discard low likelihood scores.
Claim 24
Dependent Claim
31 words
The improvement of claim 21 wherein a first threshold condition of said discarding means is based on a maximum likelihood value from a current tree search at the current time frame.
Claim 25
Dependent Claim
30 words
The improvement of claim 22 wherein a second threshold condition of said discarding means is based on an accumulated maximum likelihood score of each time frame without tree path constraint.
Claim 26
Dependent Claim
21 words
The improvement of claim 23 wherein a new node word hypothesis is accepted if said first and second conditions are satisfied.
Claim 27
Dependent Claim
29 words
The improvement of claim 24 wherein said new node word hypothesis comprises a node occurrence time and an associated likelihood score for a left and a right half-word hypothesis.
Claim 28
Dependent Claim
35 words
The improvement of claim 17 wherein said forward-backward pruning means comprises means for comparing the word hypothesis of each time frame from all vowel centers and keeping only the word hypotheses with high likelihood scores.
Claim 29
Dependent Claim
52 words
The improvement of claim 26 wherein said forward-backward pruning means operates from the first vowel center in said sequence of feature vectors to the last vowel center in said sequence of feature vectors for the forward direction and from the last vowel center to the first vowel center for the backward direction.
Claim 30
Dependent Claim
82 words
The improvement of claim 27 wherein said forward-backward pruning means operates in the forward direction including the steps: (a) calculate the maximum likelihood scores of all possible paths that end with stressed vowel for a certain word hypothesis; (b) calculate the maximum likelihood scores of all paths that end with said certain word hypothesis; and (c) sort all the path ending likelihood scores from step (b) into descending order and keep only the path ending likelihood scores that exceed a predetermined threshold.
Claim 31
Dependent Claim
66 words
The improvement of claim 16 wherein said stressed vowel center detecting means comprises means for selecting only those energy peaks represented by said feature vectors which are above a predetermined energy threshold or the feature vectors associated with the energy peaks that are acoustically matched to vowels and passing the indices of the time location of the selected energy peaks to said bidirectional tree search means.
Claim 32
Dependent Claim
84 words
An improved high-speed continuous speech-decoding system for decoding speech sentences into decoded word strings, said system including means for converting a speech utterance into feature vectors, a word hypothesizer for generating a word lattice, and means for performing a Viterbi beam search, the improvement therein comprising: said word hypothesizer including, means for representing a lexicon having a plurality of lexicon entries containing primary stressed vowels for said speech-decoding system in a vowel-centered tree which is rooted in the primary stressed-vowels of said lexicon entries.
Claim 33
Dependent Claim
84 words
An improved high-speed continuous speech-decoding system for decoding speech sentences into decoded word strings, said system including means for converting a speech utterance into feature vectors, a word hypothesizer for generating a word lattice, and means for performing a Viterbi beam search, the improvement therein comprising: said word hypothesizer including, means for representing a lexicon having a plurality of lexicon entries containing primary stressed vowels for said speech-decoding system in a vowel-centered tree which is rooted in the primary stressed-vowels of said lexicon entries.
References Cited (Backward Citations)
US4882759A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 1997-09-19 | 1998-03-21 | 1998-09-20 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2001-09-19 | 2002-03-21 | 2002-09-20 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2005-09-19 | 2006-03-21 | 2006-09-20 | $7700.00 | Window Closed / Lapsed |