This application claims the benefit of European Application No. 00127484.4, filed Nov. 29, 2000 at the European Patent Office.
1. Technical Field
The invention generally relates to the field of computer-assisted or computer-based speech recognition, and more specifically, to a method and system for improving recognition quality of a speech recognition system.
2. Description of the Related Art
Conventional speech recognition systems (SRSs), in a very simplified view, can include a database of word pronunciations linked with word spellings. Other supplementary mechanisms can be used to exploit relevant features of a language and the context of an utterance. These mechanisms can make a transcription more robust. Such elaborate mechanisms, however, will not prevent a SRS from failing to accurately recognize a spoken word when the database of words does not contain the word, or when a speaker's pronunciation of the word does not agree with the pronunciation entry in the database. Therefore, collecting and extending vocabularies is of prime importance for the improvement of SRSs.
Presently, vocabularies for SRSs are based on the analysis of large corpora of written documents. For languages where the correspondence between written and spoken language is not bijective, pronunciations have to be entered manually. This is a laborious and costly procedure.
U.S. Pat. No. 6,064,957 discloses a mechanism for improving speech recognition through text-based linguistic post-processing. Text data generated from a SRS and a corresponding true transcript of the speech recognition text data are collected and aligned by means of a text aligner. From the differences in alignment, a plurality of correction rules are generated by means of a rule generator coupled to the text aligner. The correction rules are then applied by a rule administrator to new text data generated from the SRS. The mechanism performs only a text-to-text alignment, and thus does not take the particular pronunciation of the spoken text into account. Accordingly, it needs the aforementioned rule administrator to apply the rules to new text data. The mechanism therefore cannot be executed fully automatically.
U.S. Pat. No. 6,078,885 discloses a technique which provides for verbal dictionary updates by end-users of the SRS. In particular, a user can revise the phonetic transcription of words in a phonetic dictionary, or add transcriptions for words not present in the dictionary. The method determines the phonetic transcription based on the word's spelling and the recorded preferred pronunciation, and updates the dictionary accordingly. Recognition performance is improved through the use of the updated dictionary.
The above discussed techniques, however, share the disadvantage of not being able to update a speech recognition vocabulary on large scale bodies of text with minimal technical effort and time. Accordingly, these techniques are not fully automated.
It is therefore an object of the present invention to provide method and system for improving the recognition quality and quantity of a speech recognition system. It is another object to provide such a method and system which can be executed or performed automatically. Another object is to provide a method and system for improving the recognition quality with minimum technical effort and time. It is yet another object to provide such a method and system for processing large text corpora for updating a speech recognition vocabulary.
The above objects are solved by the features of the independent claims. Other advantageous embodiments are disclosed within the dependent claims. Speech recognition can be performed on an audio realization of a spoken text to derive a hypothesis textual representation (second representation) of the audio realization. Using the recognition results, the second representation can be compared with an allegedly true textual representation (first representation), i.e. an allegedly correct transcription of the audio realization in a text format, to look for non-recognized single words. These single words then can be used to update a user-dictionary (vocabulary) or pronunciation data obtained by a training of the speech recognition.
It is noted that the true textual representation (true transcript) can be obtained in a digitized format, e.g. using known character recognition (OCR) technology. Further it has been recognized that an automation of the above mentioned mechanism can be achieved by providing a looped procedure where the entire audio realization and both the entire true textual representation and the speech-recognized hypothesis textual representation can be aligned to each other. Accordingly, the true textual representation and the hypothetic textual representation likewise can be aligned to each other. The required information concerning mis-recognized or non-recognized speech segments therefore can be used together with the alignment results in order to locate mis-recognized or non-recognized single words.
Notably, the proposed procedure of identifying isolated mis-recognized or non-recognized words in the entire realization and representation, and to correlate these words in the audio realization, advantageously makes use of an inheritance of the time information from the audio realization and the speech recognized second transcript to the true transcript. Thus, the audio signal and both transcriptions can be used to update a word database, a pronunciation database, or both.
The invention disclosed herein provides an automated vocabulary or dictionary update process. Accordingly, the invention can reduce the costs of vocabulary generation, e.g. of novel vocabulary domains. The adaptation of a speech recognition system to the idiosyncrasies of a specific speaker is currently an interactive process where the speaker has to correct mis-recognized words. The invention disclosed herein also can provide an automated technique for adapting a speech recognition system to a particular speaker.
The invention disclosed herein can provide a method and system for processing large audio or text files. Advantageously, the invention can be used with an average speaker to automatically generate complete vocabularies from the ground up or generate completely new vocabulary domains to extend an existing vocabulary of a speech recognition system.
There are shown in the drawings embodiments which are presently preferred, it being understood, however, that the invention is not limited to the precise arrangements and instrumentalities shown.
The realization 10 is first input to a speech recognition engine 50. The textual output of the speech recognition engine 50 and the representation 20 are aligned by means of an aligner 30. The aligner 30 is described in greater detail with reference to
In a first embodiment of the present invention, a selector 60 can select all one word pairs for which the representation and the transcript are different (see also
Referring to
After both texts, the time-tagged transcript generated by the SRS and the representation, have been “cleaned” or processed as described above, an optimal word alignment 140 is computed using state-of-the-art techniques as described in, for example, Dan Gusfield, “Algorithms on Strings, Trees, and Sequences”, Cambridge University Press Cambridge (1997). The output of this step is illustrated in
The original audio realization recorded by the microphone 510 together with the true transcript 520 can be provided to an aligner 550. A typical output of an aligner 30, 550 is depicted in
For the text sample shown in
In a first embodiment of the invention illustrated in
A second embodiment of the present invention, as illustrated in
Number | Date | Country | Kind |
---|---|---|---|
00127484 | Nov 2000 | EP | regional |
Number | Name | Date | Kind |
---|---|---|---|
6064957 | Brandow et al. | May 2000 | A |
6076059 | Glickman et al. | Jun 2000 | A |
6078885 | Beutnagel | Jun 2000 | A |
6466907 | Ferrieux et al. | Oct 2002 | B1 |
Number | Date | Country | |
---|---|---|---|
20020065653 A1 | May 2002 | US |