Predicting auditory confusions using a weighted Levinstein distance

• US 6,073,099 A
• Filed: 11/04/1997
• Issued: 06/06/2000
• Est. Priority Date: 11/04/1997
• Status: Expired due to Term

• Current Assignee: RPX Clearinghouse LLC (RPX Corporation)
• Original Assignee: Nortel Networks Corporation
• Inventors: Fabiani, Marc, Sabourin, Michael
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/964,023
• Time in Patent Office: 945 Days
• Field of Search: 704/254, 704/2, 704/4, 704/277, 704/231, 704/238, 704/200, 704/242, 704/241, 704/256, 704/255
• US Class Current: 704/256.6
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

A confusability tool generates a confusability cost associated with two phonemic transcriptions. The confusability cost measures the likelihood that a human or machine hearing the first word will mistakenly hear the second word. The cost calculation is based on a weighting of the Levinstein distance between the transcription pair.

Claims

1. A method of determining the likelihood of confusing a first word for a second word, comprising the steps of:

phonemically transcribing first and second orthographies of the first and second words into first and second phonemic transcriptions;

calculating a Levinstein distance between the first and second phonemic transcriptions, wherein the Levinstein distance corresponds to edit operations required to transform the first phonemic transcription into the second phonemic transcription;

obtaining a phonematic transformation weight for each edit operation of the Levinstein distance; and

summing the obtained phonemic transformation weights to generate a value indicating the likelihood of confusion between the first and second words.

2. The method of claim 1, wherein the step of obtaining the phonemic transformation weights includes the substep of determining the phonemic transformation weight based on phonological features of phonemes used for edit operations.

3. The method of claim 2, wherein the substep of determining the phonemic weight comprises using the following equation:

##EQU5## where N represents the number of phonological feature types for each phoneme, α

_i is the relative significance of feature type i, functions f_i (/x/) and f_i (/y/) represent feature value of the feature type i of phoneme /x/ and /y/, and t(f(/x/), f(/y/)) represents a transformation cost between feature values f(/x/) and f(/y/).

4. The method of claim 1, wherein the step of obtaining the phonemic transformation weights includes the substep of determining the phonemic transformation weight of a transformation between two phonemes based on Hidden Markov Models of the phonemes.

5. The method of claim 4, wherein the phonemic transformation weights obtained in the substep of determining the phonemic transformation weight of the transformation between two phonemes is calculated using the equation:

##EQU6## where n represents the number of states used to model a phoneme, v_i is a heuristically set weight for each said state used to model the phoneme, and H(x_i, y_i) represents relative entropy between probability mixtures of a state I of phoneme x and a state I of phoneme y.

6. The method of claim 1, wherein the step of summing the obtained phonemic transformation weights further includes the step of normalizing the value indicating the likelihood of confusion between the first and second words.

7. The method of claim 1, wherein the step of calculating the Levinstein distance includes the substep of calculating multiple Levinstein distances between the first and second phonemic transcriptions.

8. The method of claim 1, further including the step of modifying a lexicon based on a generated value to minimize potential confusability between words in the lexicon.

9. The method of claim 1, further including the step of using a generated value as part of a speech recognition algorithm to translate a spoken word into a computer representation of the spoken word.

10. The method of claim 1, further including the step of using a generated values to build test lexicons to be used in evaluating speech recognition algorithms.

11. A computer readable medium containing instructions executable on a computer for determining the likelihood of confusing first and second phonemic transcriptions, the instructions causing the computer to:

calculate a Levinstein distance between the first and second phonemic transcriptions, wherein the Levinstein distance corresponds to edit operations required to transform the first phonemic transcription into the second phonemic transcription;

obtain a phonemic transformation weight for each edit operation of the Levinstein distance; and

sum the obtained phonemic transformation weights to generate a value indicating the likelihood of confusion between the first and second phonemic transcriptions, that correspond to spoken words.

12. The computer readable medium of claim 11, further including instructions causing the computer to determine the phonemic transformation weight for each edit operation of the Levinstein distance based on phonological features of phonemes used for the first and second phonemic transcription.

13. The computer readable medium of claim 12, further including instructions causing the computer to determine the phonemic transformation weight as:

##EQU7## where N represents the number of phonological feature types for each phoneme, α

_i is the relative significance of feature type i, functions f_i (/x/) and f_i (/y/) represent feature value of the feature type i of phoneme /x/ and /y/, and t(f(/x/), f(/y/)) represents a transformation cost between feature values f(/x/) and f(/y/).

14. The computer readable medium of claim 11, further including instructions causing the computer to determine the phonemic transformation weight between two phonemes based on Hidden Markov Models of the phonemes.

15. The computer readable medium of claim 14, further including instructions causing the computer to determine the phonemic transformation weight based on the equation:

##EQU8## where n represents the number of states used to model a phoneme, v_i is a heuristically set weight for each said state used to model the phoneme, and H(x_i, y_i) represents relative entropy between probability mixtures of a state i of phoneme x and a state i of phoneme y.

16. The computer readable medium of claim 11, further including instructions for causing the computer to normalize the value indicating the likelihood of confusion between a first word and a second word.

17. The computer readable medium of claim 11, further including instructions causing the computer to calculate multiple Levinstein distances between the first and second phonemic transcriptions.

18. A computer system containing a process comprising:

a memory that stores a confusability tool program and a transcripted phonemic lexicon, the confusability tool program including means for calculating a Levinstein distance between first and second phonemic transcriptions from the lexicon, wherein the Levinstein distance corresponds to edit operations required to transform the first transcription, obtaining a phonemic transformation weight for each edit operation of the Levinstein distance, and summing the obtained phonemic transformation weights to generate a value indicating the likelihood of confusion between the first and second phonemic transcriptions;

a processor configured to execute the confusability tool program in the memory; and

a display that provides an ordered list of entries, each entry including the first and second transcriptions and a cost based on the value indicating the likelihood of confusion between the first and second phonemic transcriptions.

19. The computer system of claim 18, further including means for calculating cost by normalizing the minimum weighted Levinstein distance of multiple possible Levinstein distances between the first and second phonemic transcriptions.

20. The computer system of claim 18, further including means for obtaining the phonemic transformation weight based on phonological features of the first and second phonemic transcriptions.

21. The computer system of claim 18, further including means for obtaining the phonemic transformation weight based on relative entropy probability mixtures between states of Hidden Markov Models of phonemes of the first and second phonemic transcriptions.

22. A method of detecting potentially confusable orthography pairs in a lexicon, comprising the steps of:

phonemically transcribing the orthographies in the lexicon;

computing a cost for each pair of orthographies in the lexicon equal to the minimum normalized weighted Levinstein distance for the orthography pair;

comparing the cost to a threshold value and discarding the orthography pairs corresponding to cost values greater than the threshold;

sorting the thresholded values in ascending order; and

modifying the orthographies in the lexicon based on the sorted thresholded values.