Speech recognition on large lists using fragments
View Patent ↗A system and method is provided for recognizing a speech input and selecting an entry from a list of entries. The method includes recognizing a speech input. A fragment list of fragmented entries is provided and compared to the recognized speech input to generate a candidate list of best matching entries based on the comparison result. The system includes a speech recognition module, and a data base for storing the list of entries and the fragmented list. The speech recognition module may obtain the fragmented list from the data base and store a candidate list of best matching entries in memory. A display may also be provided to allow the user to select from a list of best matching entries.
1. A speech recognition method in which an entry corresponding to a speech input is selected from a list of entries, the method comprising:
detecting the speech input;
recognizing a phoneme sequence of the speech input;
providing a list of fragments of entries in a list of entries, the fragments being based on a subword or phoneme level; and
comparing the phoneme sequence of the recognized speech input to the list of fragments to generate a candidate list of best matching entries based on comparison scores,
wherein a comparison score is calculated for a fragment when the recognized speech input is compared to the fragment, the comparison score being a measure of how well the recognized speech input fits to the fragment, wherein a score for one list entry is calculated based on the comparison scores of all the fragments that build the list entry,
where the fragment is accompanied by different wildcards, each wildcard representing the part of the list entry not considered in the fragment of the list entry and each wildcard having a different weight when the recognized speech input is compared to the fragment.
2. The method of claim 1 , where at least one fragment is provided for each entry of the list.
3. The method of claim 1 further comprising:
providing a list of fragments containing substantially all different fragments of the entries, where for generating the candidate list, the recognized speech input is compared to the list of fragments.
4. The method of claim 3 further comprising calculating a score for each fragment of the list of fragments.
5. The method of claim 1 further comprising calculating a score for one entry by adding the scores of the fragments that form the entry.
6. The method of claim 1 further comprising adding wildcards to one side of the fragment, to the other side of the fragment or to both sides of the fragment.
7. The method of claim 1 further comprising selecting the wildcard in such a way that the wildcard matches substantially uniformly against any recognized speech input.
8. The method of claim 1 further comprising:
when the wildcard is provided only on one side of the fragment, determining:
on which side of the fragment the wildcard is not provided,
where the corresponding side of the recognized speech input is determined, where the matching of the recognized speech input to the fragment is substantially carried out based on a part of the recognized speech input provided on said corresponding side of the recognized speech input.
9. The method of claim 1 further comprising:
when the wildcard is only provided on one side of the fragment, determining the corresponding side of the recognized speech input; and
applying a pruning mechanism to the other side of the recognized speech input.
10. The method of claim 1 further comprising:
normalizing the scores calculated when matching the recognized speech input to the list of fragments; and
adding the normalized scores to calculate the score of each entry.
11. The method of claim 1 further comprising:
calculating a neutral score by matching the at least one wildcard versus the recognized speech input, the neutral score being subtracted from all scores calculated for the different fragments.
12. The method of claim 1 further comprising providing the fragments in a tree like structure.
13. The method of claim 12 further comprising adding the scores for the different fragments building the list of entries on the basis of the tree like structure of the fragments.
14. The method of claim 1 further comprising calculating position based scores for the fragments, in which the score for a fragment depends on the position of the fragment in the entry.
15. The method of claim 1 further comprising using the candidate list of best matching entries determined on the basis of fragments of the entries as a basis for a recognition step.
16. The method of claim 1 , where the entries include at least one of the following information:
personal names, addresses, city names, street names, points of interest, song titles.
17. The method of claim 1 further comprising generating the list of best matching entries by sorting the entries according to the calculated score, the n entries having the best score building the candidate list.
18. The method of claim 1 further comprising using a context sensitive Levenshtein algorithm to match the recognized speech input to the fragments of the entries.
19. The method of claim 11 further comprising calculating a fragment specific neutral score to calculate the neutral score, the score for the entries being calculated on the basis of the fragment specific scores.
20. The method of claim 1 further comprising comparing the recognized speech input to at least
some of the complete entries in order to generate the candidate list.
21. The method of claim 1 further comprising fragmenting the entries into fragments on a sub-word
unit level.
22. The method of claim 1 , where the entry comprises a combination of a city name and a street name and a fragmentation into fragments is at least carried out between city name and street
name.
23. The method of claim 1 further comprising selecting the fragments of the entries in such way that the fragments for one entry overlap.
24. The method of claim 1 further comprising selecting the fragments of the entries in such a way that the fragments cover only a part of the entry.
25. The method of claim 1 further comprising storing the entries together with wildcards, where the wildcards are indicative that a user's utterance for selecting one entry from the list of entries contains more than the entry itself.
26. A speech recognition system in which a speech input is used to select an entry from a list of entries, the system comprising:
a data base having a list of entries and a list of fragments, each fragment representing a part of one entry; and
a hardware-based speech recognition module configured to recognize a speech input and to compare the recognized speech input to the list of entries in order to generate a candidate list of best matching entries based on the comparison result, where the speech recognition module compares the recognized speech input to the list of fragments in order to generate the candidate list
where the fragment is accompanied by different wildcards, each wildcard representing the part of the list entry not considered in the fragment of the list entry and each wildcard having a different weight when the recohnized speech input is compared to the fragment.
27. A non-transitory computer readable medium having logic configured for execution in a speech recognition system, the logic comprising:
logic for detecting the speech input;
logic for recognizing a phoneme sequence of the speech input;
logic for providing a list of fragments of entries in a list of entries, the fragments being based on a subword or phoneme level; and
logic for comparing the phoneme sequence of the recognized speech input to the list of fragments to generate candidate list of best matching entries
wherein a comparison score is calculated for a fragment when the recognized speech input is compared to the fragment, the comparison score being a measure of how well the recognized speech input fits to the fragment, wherein a score for one list entry is calculated based on the comparison scores of all fragments that build the list entry and where the fragment is accompanied by different wildcards, each wildcard representing the part of the list entry not considered in the fragment of the list entry and each wildcard having a different weight when the recognized speech input is compared to the fragment.