Method and Apparatus for Text-Independent Speaker Recognition

This invention relates to the general field of voice processing and more particularly to a method and apparatus for recognition of a speaker independent of the text being spoken.

Previous text-independent speaker recognition systems have used long term averages of various speech parameters to model a talker's voice. These techniques therefore require large amounts of speech for both modeling and recognition in order to estimate accurately the mean and variance of the speech parameters. In addition, the systems are very sensitive to channel variations that distort the long term average. Both these limitations have restricted the development of text-independent speaker recognition systems for many tactical applications.

The present invention overcomes the above-mentioned problems by using individual short duration speech frames from a speaker as models of that speaker's voice. Each frame characterizes a particular speech production of the speaker. Model frames are discarded that represent speech events that do not occur frequently in the modeling data. Recognition is based on frame by frame measurements rather than long term averages, so that large amounts of speech are not required. Input frames from the unknown speech whose speaker is to be recognized which are not close to any of the model frames are ignored. In this way reliable recognition can be achieved even if new speech events that did not occur in the model data are encountered in the speech data being recognized. It is also an advantage that the modeling technique above described is much less sensitive than previous techniques to changes in the channel over which the speech is obtained.

Above-mentioned and other features and objects of this invention will become more apparent by reference to the following description taken in conjunction with the accompanying drawings, in which:

FIG. 1 is a block diagram of the preferred embodiment speaker recognition apparatus.

FIG. 2 is a block diagram showing the steps of the preferred method of carrying out the present invention.

FIG. 3 is a more detailed block diagram of the step of generating a character set shown in FIG. 2.

FIG. 4 is a more detailed block diagram of the step of generating character frequency of occurrence data shown in FIG. 2.

FIG. 5 is a more detailed block diagram of the step of generating character occurrence statistics shown in FIG. 2.

FIG. 6 is a more detailed block diagram of the step of selecting model characters shown in FIG. 2.

FIG. 7 is a more detailed block diagram of the step of calculating distance statistics shown in FIG. 2.

FIG. 8 is a more detailed block diagram of the final step of recognition shown in FIG. 2.

FIG. 9 is a curve of the telephone channel simulator 1000 shown in FIG. 10.

FIG. 10 is an alternate block diagram of the invention of FIG. 1 showing model speech data and unknown speech data being received over separate channels.

FIG. 1 is a block diagram of the preferred embodiment text-independent speaker recognizer apparatus designated generally 100. The invention performs recognition in two steps: first, digitized and processed samples (frames) of speech for each speaker candidate (in general, the apparatus and method is capable of recognizing a plurality of speakers, e.g. 11) are selected for use as speaker models which models are then stored; and, second, each frame of speech from an unknown speaker to be recognized is compared with each frame in the stored speaker models to recognize the unknown speaker. An important part of the recognition process is in determining which speech samples to use as speaker models.

In the preferred embodiment, the data used to generate the speaker models comprises two minutes of speech (a speaker data file) from each of eleven speaker candidates, with each two minutes being divided into 12 ten second segments of speech. Apparatus 100 comprises speech converting means designated generally 102 connected by bus 120 to general processor designated 140. General processor 140 is coupled to external memory 160 and special processor 170. Speech enters apparatus 100 as audio signals through the microphone 104 where it is subsequently amplified by mike preamp 106 of conventional design. Next, the audio is sampled by conventional sample and hold circuitry 108 8000 times per second in response to signals from an 8 kHz sample clock 110. The output of sample and hold circuit 108 is analog to digital (A/D) converted by A/D circuit 112 (again in response to signals from 8 kHz clock 110) to produce 16 bit pulse code modulated (PCM) signal samples. The PCM signals are transmitted via bus 120 to general purpose processor 140 where they are double buffered in 16 kilobyte buffer memories 142 and 144. Approximately one second of 8 kHz 16 bit PCM samples can be stored in a 16 kilobyte buffer memory. The speech samples are then transmitted to external memory 160 where they are stored for further processing in the special processor 170. A suitable memory is an RM05 300 megabyte disk drive.

The first step in generating speaker model data is to convert the raw PCM data into frames of speech data. This is done by taking approximately 200 sequential PCM samples and, using well known linear predictive coding (LPC) techniques to model the human vocal tract (See "Linear Prediction: A Tutorial Review", by J. Makhoul, Proceedings of IEEE, April 1975, Volume 63 No. 4, pages 561-580; and "Linear Prediction of Speech", by J. P. Markel and A. H. Gray, Jr., Springer-Verbg, 1976), converting these samples into an LPC twelve coefficient frame of speech. Them taking the last 100 PCM samples in time from the first conversion and combining it with the next 100 PCM samples in time, a new LPC-12 frame is formed. Each LPC-12 frame is a point in a multidimensional speaker space which is a speaker spacer which models the speaker's vocal tract. Approximately, one second of data is transferred at a time from memory 160 via 16 kilobyte buffer memory 146 for processing in special processor 170. In the preferred embodiment special processor 170 is a Floating Point System AP120B Array Processor equipped with 4K words of its own program memory for storing the LPC program and the model generation and recognition programs and 64K 38 bit words of data memory. Each speaker data file is processed as above described to generate speaker model data comprising LPC frames for each speaker. The data is then stored in the external memory 160.

Referring now to FIG. 2, a block diagram flow chart 200 of the process of recognizing a speaker from the digitized speaker data files is presented. Firstly, a character set is generated 202 by discarding many of the frames of data selecting a finite number of them as representative of the speaker field. The selected frames known as a character set are selected from the speaker model data from all eleven speakers by using a character covering algorithm which insures that no two characters selected as representative of the speaker field will be separated by a distance less than a predetermined amount, T. One suitable method for determining the distance between characters in the speaker field is the well known Itakura distance metric. In the preferred embodiment, T equals 1.5 Itakura units. Further details on generating a character set are shown in FIG. 3. Speaker model data for one of the speakers is inputted 302 from external memory 160. Characters are generated frame by frame 304. The first frame of the first speaker's model data becomes a character and then a new frame is processed 306 and 308. The distance from each new frame to all previously selected characters is determined 310 until either D is less than T in which case the frame is discarded and a new frame processed or the determination is made that the current frame should be a new character since the distances from the frame to all previous characters are greater than or equal to T 312, 314 and 316. All the frames for each speaker's model data are processed in this way 318 and 320. The result is a character set representative of the speaker field for the speakers chosen in which no two frames are closer than T. In the preferred embodiment apparatus and method of the present invention, this results in a character set numbering typically 1000 characters.

Referring now to FIGS. 2 and 4, the next step in speaker recognition after generating a character set for all eleven speakers is to generate character frequency of occurrence data 204 for each speaker's model data separately. Details of this step are given in FIG. 4. The number of times each character 402 in the character set 404 is the closest character 406 to the current frame 408 of a 10 second segment of speech 410 is determined 412 for each previously selected speaker data file. The character occurrence data is normalized by the number of frames processed within each 10 second segment 414, 416. When all the segments have been processed for a selected speaker data file, a new speaker data file is inputted from remote memory 160 into the array processor 170 along with the character set for processing. The result of this is character occurrence data for each of 12 10 second segments for each individual speaker.

Referring now to FIGS. 2 and 5, the character occurrence data is processed to generate character occurrence statistics 206. This step is shown in more detail in FIG. 5. All of the normalized character occurrence data is read into the general purpose processor 140 from the disk drive 160, step 502 in FIG. 5. One of the eleven speakers is chosen as speaker 504 and the remaining speakers become collectively the imposters. Then for each character in the character set 506 the mean frequency of occurrence of that character across all 10 second segments of the speaker's data file is determined 508. The normalized frquency of occurrence data is used for this purpose. For example, suppose character No. 1 was the character with minimum distance to 8 of the frames in the first 10 second segment, there being approximately 800 frames in each 10 second segment. Then the normalized character occurrence for character No. 1 for that 10 second segment is 0.01. Similarly, character No. 1 would have values for each of the eleven other 10 second segments for the speaker data file. The mean frequency of occurrence would be the sum of normalized character occurrence data for character No. 1 for the speaker divided by 12, the number of 10 second segments in the speaker's data file. Next, using the mean value just calculated, the speaker standard deviation is calculated 510 in a conventional manner. Finally, for character No. 1, the mean frquency of occurrence across all 10 second segments for the imposter (all speaker data files not chosen as the speaker) is calculated from the normalized character occurrence data for character No. 1 512. For example,.the sum would be divided by 120 since there are ten remaining speakers, each with 12 10 second segments. Using the imposter mean, the imposter standard deviation is calculated 514 as above. The speaker and imposter mean and standard deviation for each character in the set 516 is determined in similar manner. For each character the speaker and imposter mean and standard deviation are saved 518. When all characters have been processed, a new one of eleven speaker data files is chosen and the remaining 10 become collectively the imposter and the above process is repeated 520. This is repeated until each of the speaker data files has been treated as the speaker with the remaining data files collectively being the imposter.

It should be remembered that in the preferred embodiment the character set used to represent the speaker field for the eleven speakers contained approximately 1000 characters. These 1000 characters are not particularly identified with any one of the eleven speakers. Using the character occurrence statistics generated by steps 204 and 206, a model set of characters for each speaker is selected which distinguishes that speaker from the remaining speakers 208. Referring to FIG. 6, this step is shown in more detail. All of the speaker and imposter character means and standard deviations data is inputted into general processor 140, see step 602. One of the speakers is selected as speaker and then a character from the character set of 1000 characters is selected 604. A parameter f is set equal to zero 606 and the speaker mean and imposter mean for the selected character is examined. If the speaker mean is greater than 0.9 times the imposter mean then f is set equal to ##EQU1## See step 608. This process is repeated for each of the 1000 characters in the character set 610. Then those characters with the forty highest f values are chosen as the speaker's model characters 612 and 614. Then a new speaker is selected and new model characters are generated until all speakers have their own model characters.

A final step in generating a speaker model before recognition is effected is to generate distance statistics for an output model for each speaker 210. This final step in model generation is shown in detail in FIG. 7. All speaker character files are loaded into array processor 170, see step 702. Then one of eleven speaker data files is selected and loaded into array processor 704. An input frame 708 of a selected ten second segment 706 is chosen and the distances, D, from the input frame to the closest character in each of the eleven speakers' model character sets are determined and saved in a model histogram, one for each set of model characters 709, 710, 712 and 714. The remaining frames in the selected segment are similarly processed 715. (A model histogram is a frequency distribution of the distances D measured.)

For each completed segment, for each model histogram, the distance, "D", such that 30% of the frames had distances less than "D" and 70% of the frames had distances greater than "D" is determined 716. The values "D" for each character set for each segment are saved 718. When all of the segments for the selected speaker data file are completed 720 the speaker whose data file was processed becomes the speaker and the remaining speakers are imposters. Then the mean distance value of "D" and variance of "D" across all speaker segments is calculated and then the mean distance value and variance across the imposters are calculated 724. These values are appended to the selected speaker's set of model characters to form a speaker model. The above processing is repeated for each of the speaker data files in turn to form eleven speaker models.

This completes the model generation phase 201 of the speaker recognition process. The final step of recognition of an unknown speaker 212 is shown in more detail in FIG. 8. All of the speaker models (40 characters plus statistics per model) are loaded into the array processor 170, see step 802. As the speaker to be recognized is speaking, a frame from the unknown speech is selected 804 and the distance from the frame to the closest character for each speaker model is determined and saved in a model histogram 806, 808, 810 and 812, there being a histogram for each speaker model. When all of the frames from the available segment (unknown length possibly less than ten seconds) of unknown speech have been processed as above 814, one of the eleven speaker models is selected as speaker 816 and the distance value "D" for that model's histgram such that 30% of the frames had distances less than "D" and 70% of the frames had distances greater than "D" is determined 818. Then, using the speaker and imposter statistics (mean values and variances contained in the speaker model), the probability that the selected speaker (step 816) would produce the distance "D" just determined, Prob (D/Spk), and the probability that an imposter would produce the distance "D", Prob (D/Imp), are calculated in conventional manner from first principles of probability theory 820 assuming Gaussianly distributed distances. Next, the likelihood ratio Prob (D/Spk)/Prob (D/Imp) is formed and saved 822. Then a new one of the eleven speaker models is chosen as speaker and a new likelihood ratio is determined as described above. This is repeated until each of the eleven speaker models has served as the speaker.

The likelihood ratios are transmitted via bus 172 to the general processor 140 (VAX 11/780) where the speaker with the highest likelihood ratio is selected 826. In the embodiment shown in FIG. 1, the selected speaker's identification or name is then displayed on a CRT 174.

It should be remembered that one suitable method for determining the distance between one frame (character) and another is the Itakura method for determining the distance metric. This method can be used wherever a distance is to be determined as described above in connection with a description of the Figures.

The above described apparatus and method is expected to provide short utterance speaker recognition capability, i.e., within a one second delay of the utterance a speaker can be recognized where the utterance lasts approximately ten seconds. Furthermore, it was demonstrated that high recognition accuracies (approximately 96%) can be obtained even where the speaker data used for the models and the test data to be recognized were recorded one week apart (as long as the channels over which the speech data, both model and test data, are communicated are approximately the same for both).

The above described method and apparatus works well even where the channel for the unknown differs from that of the model data. For example, speaker data files were processed with a channel simulator to model passing the data through a typical telephone channel. The bandwidth of the channel was approximately 500-3000 Hz with about seven db of ripple in the pass band. See FIG. 9. As FIG. 10 shows, the unknown speech input passes through the telephone channel 1000 before entering the apparatus while the model data enters 102 directly. An added processing step, blind deconvolution 1002, is added to the PCM samples before LPC-10 conversion in the Array Processor. This is a well known technique employed in telephony to provide channel normalization. However, in so doing in prior art speaker recognition systems the blind deconvolution also tends to normalize speaker characteristics making the prior art speaker recognition techniques useless or much less effective. Test results of the present invention indicate that for a 10 second long unknown utterance, recognition accuracy only drops from 96% where there is no channel differences between model data and test data recordings to 89% where channel differences were present and blind deconvolution was employed.

The source code for the program implementation of the steps described herein is provided on Pages 13-47. The source code is made up of four different languages: for the programs executed in the VAX processor 140 (those listings having a title ending in XXXX.C) the source code is written in the "C" language; for the programs executed in the array processor there are two types:

(a) source code written in Array Processor Assembly Language (those listings having a title ending in XXXX.APS), and

(b) source code written in "C-Vector Function Chainer Language" (those listings having a title ending in XXXX.V); and make-file code listings. All array processor code was compiled and assembled using a software package available from APUNIX, a software company located in San Diego, California.

While the present invention has been disclosed in connection with the preferred embodiment thereof, it should be understood that there may be other embodiments which fall within the spirit and scope of the invention as defined by the following claims. ##SPC1##