Method of and apparatus for deriving a plurality of sequences of words from a speech signal

• US 5,987,409 A
• Filed: 09/26/1997
• Issued: 11/16/1999
• Est. Priority Date: 09/27/1996
• Status: Expired due to Fees

• Current Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Original Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Inventors: Steinbiss, Volker, Seide, Frank, Tran, Bach-Hiep
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/938,922
• Time in Patent Office: 781 Days
• Field of Search: 704/240, 704/256, 704/255, 704/251, 704/231, 704/253, 704/257
• US Class Current: 704/240
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

The determination of a plurality of sequences of words from a speech signal with a decreasing probability of correspondence utilizes the best word sequence as a basis and as further word sequences there are determined only those which enclose a part of the best word sequence, that is to say the remainder of these word sequences. To this end, the recognition involves first the formation of a word graph and the best word sequence is separately stored as a tree which initially has one branch only. The word boundaries of this word sequence form nodes in this tree. Because only nodes of this tree have to be taken into account for the next-best word sequences, the calculation is substantially simpler than if the complete word graph were first completely expanded in the form of a tree and completely searched again for each new word sequence.

Claims

1. A method of deriving at least one sequence of words of a predetermined vocabulary from a speech signal, including the following steps:

a) providing a store of words and reference signals, different words being associated with different sequences of said reference signals,

b) deriving test signals from the speech signal,

c) comparing the test signals with said different sequences of said reference signals,

d) determining words with scores based on the comparisons of test signals and forming a word graph therefrom which contains nodes as word boundaries and words as edges,

e) identifying a first path through the word graph with the smallest sum of scores of the sequence of words,

characterized in that the method further comrpises the steps of;

f) storing, in order to output further sequences of words having a next-larger sum of scores, the sequence of words on the first path as a basis for a tree having tree nodes at the word boundaries and one end node,

g) forming, for each tree node and for the path segment of the word graph which extends to the corresponding graph node, the overall sum of scores which are situated thereon as from the start of the word graph, the path segments with the sequences of words determined thus far being excluded therefrom,

h) identifying the sequence of words for which the sum of the scores of the words on a path segment in the word graph as far as the graph node corresponding to a tree node, and from this tree node in the tree, is smallest, the already stored tree being extended by the sequence of words on this path segment, and

i) repeating steps g) and h) at least once, and

j) outputting a plurality of mutually different sequences of words having progressively higher scores.

2. A method as claimed in claim 1, characterized in that the method further comprises:

storing, for each respective tree node, the sum of the scores of the words on the best path segment in the word graph prior to the graph node corresponding to said respective tree node, and from said respective tree node to the end node in the tree,

outputting the sequence of words associated with the minimum sum of the scores of all nodes as the further sequence,

marking the tree nodes and associated sums of scores, associated with the path segment with the outputted words,

determining and storing the sum of the scores for the marked tree nodes only, and

determining a next further sequence based on the sum of the scores for the marked tree nodes.

3. A method as claimed in claim 1, characterized in that for each path values indicating the probability of sequences of n≧

1 given words are taken into account.

4. A method as claimed in claim 3, characterized in that the method further comprises:

storing, for each respective tree node, the sum of the scores of the words on the best path segment in the word graph prior to the graph node corresponding to said respective tree node, and from said respective tree node to the end node in the tree,

outputting the sequence of words associated with the minimum sum of the scores of all nodes as the further sequence,

marking the tree nodes and associated sums of scores, associated with the path segment with the outputted words,

determining and storing the sum of the scores for the marked tree nodes only, and

determining a next further sequence based on the sum of the scores for the marked tree nodes.

5. An apparatus for deriving a plurality of sequences of words of a predetermined vocabulary from a speech signal, comprising:

a sequence and acoustic recognition system for deriving a word graph based on the speech signal,

means for assigning probability scores for each word in the word graph,

means for storing and outputting at least one word sequence, based on said word graph, having a minimum sum of scores,

a memory for storing the output word sequence in the form of a tree,

a processor, including means for receiving the word graph, for deriving from the word graph, after the first word sequence output, only those word sequences having at least one word boundary in common with the already stored word sequences, and excluding sub-sequences of words which start from the beginning of the word graph and at least the last word of which corresponds to a sub-sequence stored in the memory, and storing in said memory from each newly output word sequence only the parts which differ from the word sequences already stored, and

means for additionally outputting said word sequences having at least one word boundary in common with the already stored word sequences and excluding said sub-sequences of words.