KR20040107173A - An Amplitude Warping Approach to Intra-Speaker Normalization for Speech Recognition - Google Patents

Abstract

PURPOSE: An amplitude varying method for intra-speaker normalization in speech recognition is provided to achieve amplitude normalization according to pitch-varied pronunciation and compensate an amplitude difference by estimating a warping factor in a speaker, to thereby improve a recognition rate and reliability of speech recognition applications. CONSTITUTION: A reverse ratio between an input pitch and a reference pitch of a speaker is calculated to decide an amplitude warping factor. The height of a triangle filter is controlled while the amplitude is controlled using the amplitude warping factor and, simultaneously, the gradient of the amplitude on the frequency axis is determined. After the gradient of the amplitude is determined, the result feature vectors are HMM(Hidden Markov Model)-decoded.

Description

An Amplitude Warping Approach to Intra-Speaker Normalization for Speech Recognition}

The present invention relates to a method for changing the amplitude for in-speaker normalization in speech recognition, and more particularly, to achieve amplitude normalization according to pitch change utterance by estimating a new intra-warping factor for speaker normalization. Amplitude for in-speaker normalization in speech recognition, which can improve recognition rate by applying recognition model appropriate to user's voice in application or speaker adaptation method to recognizer. It is about a change method.

The vocal cords of the vocal cords are commonly known to control the pitch of voices, while the saints are known to form vowels and form consonants through formants. Formant components are almost independent of the speech signal.

In an effort to reduce speech recognition performance due to vocal morphology between speakers, frequency warping techniques have been studied for speaker normalization. That is, techniques for normalizing parametric component representations of speech signals have been studied to reduce the effects of speaker-to-speaker differences.

Thus, normalization was performed using linear and nonlinear frequency warping functions to compensate for formant position variations between speakers. These procedures are consistent with each speaker's actual saint shape and attempt to solve the complex problem of formant position estimation according to the compensation for these differences.

When using Gaussian mixtures as output distributions in the Hidden Markov Model, one of the most important problems is that the various speaker-dependent scale factors tend to be modeled by the polymorphism of the mix distributions. have.

In addition, the inter-speaker element plays an important role in speech recognition, and is based on Vocal Tract Normalization for speaker adaptation that requires inter-speaker normalization. This is attempted to reduce the fluctuation of voice between speakers by variation compensation in pitch change speech according to the state of emotion.

The existing saint normalization method is a very good way to improve the accuracy of the speaker-to-speaker normalization, based on the frequency axis normalization for speaker-to-speaker normalization.

Here, the frequency axis normalization for speaker-to-speaker normalization is as follows.

The most important idea of VTN is to normalize the frequency axis of acoustic vectors for each speaker to remove speaker dependent variability in acoustic vectors during recognition. The positions of the formant vertices of the spectrum for a given sound utterance are inversely proportional to the length of the saints.

In this case, the length of the saints varies from about 13cm to 18cm, and the formant center frequency may vary by about 25% among the speakers. This means that the formant center frequency varies from person to person and varies by about 25% due to differences in saints' lengths.

These variance factors are major deterioration factors of speaker-dependent to speaker-independent speech recognition performance.

The optimal warping factor is obtained by searching for 13 elements of equal intervals between 0.88 ≦ α ≦ 1.12. For example, dividing alpha by 0.02 intervals in the above range yields 13 elements: 0.88, 0.90, 0.92, ..., 1.12, and so on.

The range of α is chosen to reflect a change of approximately 25% range in vocal lengths found in adults.

Several methods have been suggested to determine the optimal frequency warping magnitude in recognition.

In speech recognition, observe the sequence of acoustic vectors over time t = 1, ..., T. In other words,

For each hypothesized word sequence W, we assume a model distribution p (X | W; θ) with appropriate reference model parameters θ. In speaker-adaptive acoustic modeling, it is assumed that the following dependence on the distribution of the characteristic parameter α of a speaker is assumed.

p (X | W; θ, α)

Typically, you distinguish between two variables.

First, the model as the transformation of the parameter θ, and converts it into a parameter θ ^α normalizing the non-normalized model parameters θ for each value of the speaker between the characteristic parameter α.

θ → θ ^α

Therefore, the distribution becomes p (X | W; θ, α) = p (X | W; θ ^α ) .

Second, as a transform of observation vector X, this can be formulated as a mapping of acoustic vectors.

X → X ^α

At this time, the distribution becomes p (X | W; θ, α) = p (X ^α | W; θ) .

As described above, the conventional vocal tract normalization (VTN) method compensates for the change in the frequency axis of the spectral envelope component of vocalization caused by different vocal lengths from person to person through the normalization of the frequency axis. Is compensated, but the amplitude difference is not compensated.

Therefore, in view of the fact that the existing vocalization normalization method compensates only for the difference in frequency, the present invention achieves amplitude normalization according to pitch change utterance and estimates amplitude through a new intra-warping factor estimation for speaker normalization. In order to compensate for the difference, the recognition model suitable for the user's voice is applied to the speech recognition application product or the speaker adaptation technique is applied to the recognizer to improve the recognition rate and improve the reliability of the speech recognition application product. The purpose of this paper is to provide a method for changing the amplitude for in-speaker normalization.

1 and 2 show linear prediction coefficient (LPC) spectral envelopes of voiced sounds produced by males and females according to pitch change utterances.

Fig. 3 shows voice spectra produced by the normal (bold line) vocalization of male voice / a / and the lowered pitch (dotted line).

4 shows a specific parameter β in the speaker according to the pitch change speech.

5A and 5B show a mixture-based optimal frequency warping factor estimation.

Figure 6 shows a mel-filter bank analysis with frequency warping.

7 shows a mel-filter bank with amplitude warping for speaker normalization.

8 shows a method of applying α and β in order.

The present invention for achieving the above object is:

The specific spatial distributions of speech that are not deformed from the speaker's pitch-changing speech vary due to the acoustical differences of speech generated by voice and vocal tracts, so that the amplitude is calculated by calculating the inverse ratio between the input and reference pitches of the input speaker. Determine the warping factor; In the speech recognition, the amplitude change method for normalization in a speaker is characterized in that the amplitude of the triangular filter is adjusted using the amplitude warping factor, and the slope of the amplitude is determined on the entire frequency axis.

In a preferred embodiment, the calculation of the inverse ratio between the input pitch and the reference pitch of the input speaker is given by p1 for the normal (average) speech and p2 for the new voice. In general, if p2 is high, the amplitude is also high, so that it is calculated by adopting p1 / p2 as a new gradient control constant value to adjust the amplitude to the amplitude of p1 for normalization.

In a more preferred embodiment, after the slope of the amplitude is determined, the resulting feature vectors are HMM decoated.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

First, the process used to perform the amplitude warping approach to in-speaker normalization is described in detail. These processes are attempts to reduce fluctuations in voice in the speaker by compensating for fluctuations in pitch change utterances caused by emotion.

Since distortions due to pitch change vocals can be designed by simple linear warping in the frequency domain of the speech signal, the normalization procedure adjusts the amplitude axis by an appropriately estimated warping factor.

Rhymes is commonly known as the expression of acoustical characteristics of emotions. The feature parameter is analyzed from the voiced sound region of the speech waveform data. This is an important point for the speaker element.

1 and 2 show linear prediction coefficient (LPC) spectral envelopes of voiced sounds produced by males and females according to pitch change utterances.

That is, Figure 1 is a male voiced LPC spectrum envelopes (vowel / a / with a pitch of 113 ~ 251 Hz), Figure 2 is a female voiced LPC spectrum envelopes (194 ~ 342 Hz having a local pitch of As the vowel / a /), the reason for the energy gain at higher harmonics is shown by a comparison of the globular airflow waveform. As the intensity of pronunciation increases, the closure rate of the gate increases. In general, male voices tend to have lower fundamental frequencies and stronger harmonics than female voices.

FIG. 3 shows voice spectra produced by the normal (bold line) vocalization of male voice / a / and the lowered pitch (dashed line). The characteristic spatial distributions of speech that are not deformed from the speaker's pitch-changing speech vary due to the acoustical differences caused by voice and vocalization. Therefore, it is possible according to the present invention to consider the warping factor by inverse calculation of the reference pitch and the input pitch.

More specifically, the warping factor represents the inverse ratio between the pitch of the input speaker and the reference pitch.

Accordingly, the slope of the amplitude on the entire frequency axis is easily determined while controlling the amplitude while adjusting the height of the triangular filter using the amplitude warping factor.

As shown in FIGS. 6, 7, and 8, the MFCC method analyzes the speech by applying a triangular filter to the frequency axis. In other words, the triangular filter is used when the frequency axis is divided into predetermined bandwidths.

At this time, if the pitch (reference pitch) at normal (average) speech is set to p1 and the pitch at the time of newly speaking the same sound is set to p2, the amplitude at p1 is normalized because the amplitude is generally higher when p2 is high. To adjust the degree, p1 / p2 is adopted as the new slope adjustment constant value.

The resulting feature vectors are then used for HMM decoding, and the goal is to warp the amplitude scale of each test vocalization to a normalized HMM model.

4 shows the characteristic parameter β in the speaker according to the pitch change utterance. The characteristic scale factor in the speaker is closely related to the energy of the spectrum. 4 uses pitch and energy Estimate β satisfying.

In the normalization of acoustic modeling in a speaker, it is assumed that it depends on the distribution of characteristic parameters β in a speaker.

p (X | W; θ, α, β)

Typically you will distinguish between two transformations.

First, the model as the transformation of the parameter θ, is converted into a model parameter θ ^β normalizing the non-normalized model parameters with respect to the speaker within the characteristic parameter β value.

θ → θ ^{α, β} (or θ → θ ^β )

The distribution is as follows.

p (X | W; θ, α, β) = p (X | W; θ ^{α, β} )

or

p (X | W; θ, α, β) = p (X ^α | W; θ ^β )

Second, as a transformation of the observation vector X, it can be formulated according to the mapping of acoustic vectors.

X → X ^{α, β} (or X → X ^β )

The distribution is as follows.

p (X | W; θ, α, β) = p (X ^{α, β} | W; θ)

or

p (X | W; θ, α, β) = p (X ^β | W; θ ^α )

The scale element β in the speaker of the amplitude axis is used to rescale the amplitude axis before calculating the acoustic vector for speech recognition.

In the speech recognition according to the present invention, the amplitude change method for in-speaker normalization was tested using SKKU-1 (SKKU: SungKyunKwan University) voice database (DB).

The vocabulary of SKKU-1 voice DB is composed of Korean digits, name, phonetically balanced word (PBW) and phonetically rich word (PRW).

The voice signal was emphasized as high as 1-0.95z ^-1 and the Hamming window of 20ms was taken and analyzed in units of 10ms. Each frame extracted 39-dimensional feature vectors.

The features are the 12th-order mel-frequency cepstrm coefficient (MFCC) vector, the 12th-order delta-MFCC vector, the 12th-order delta-delta-MFCC vector, log energy, delta log energy, and delta-delta log energy.

On the other hand, when the phonetically balanced word (PBW) selects a phonologically balanced word, the phonetically rich word (PRW) includes more words than PBW and includes thousands of words and there is no provision for word count. The Hamming window is a representative window function used in a unit analysis section when analyzing a voice.

5A shows an estimate of an optimal frequency warping factor based on a mixture. 5B shows an estimate of the optimal frequency warping factor based on the frequency warping of the input speech. Negative was warped using the estimated warping factor, and the results of the feature vectors are used for HMM decoding.

Figure 6 shows a mel-filter bank analysis with warping. For amplitude normalization, first, we extract pitch and energy from vocalization. And secondly we determine the parameters in the speaker.

7 shows a mel-filter bank with amplitude warping for speaker normalization.

More specifically, FIG. 6 shows a method for analyzing when the factor α is determined, FIG. 7 shows a method for analyzing when β is determined, and FIG. 8 shows a method of applying α and β in order.

Table 1 below shows the recognition words of numbers and words in SKKU-1 DB using the base recognizer, the base base recognizer with inter-speaker normalization, and the base base recognizer with inter-speaker and intra-speaker normalization. Show error rate.

As shown in Table 1 above, the baseline is when normalization is not applied, with α is when only existing frequency normalization is applied, and with α and β is when conventional frequency normalization and proposed amplitude normalization are applied. As the recognition error rate is shown, the error decreases gradually.

That is, according to the recognition results, the word recognition rates were 96.4% and 98.2%, and the error rate was reduced from 0.4 to 2.3% for numbers and word recognition.

As described above, according to the method of changing the amplitude for in-speaker normalization in speech recognition according to the present invention, amplitude normalization according to pitch change utterance can be achieved through a new in-speaker warping factor estimation for speaker normalization. have.

Therefore, it is possible to improve the recognition rate and improve the reliability of the speech recognition application product by applying a recognition model suitable for the user's voice in the speech recognition application product or by applying the speaker adaptation technique to the recognizer.

Claims (3)

Characteristic spatial distributions of voices that are not deformed from the speaker's pitch-changing voices vary due to the acoustical differences of voices generated by voice and vocal tracts, so that the amplitude is calculated by calculating the inverse ratio between the input pitch and the reference pitch Determine the warping factor; The amplitude changing method for normalization in the speaker in the speech recognition, characterized in that while controlling the amplitude while adjusting the height of the triangular filter using the amplitude warping factor, the slope of the amplitude on the entire frequency axis. The method of claim 1, wherein the calculation of the inverse ratio between the input pitch and the reference pitch of the input speaker is performed by setting the pitch (reference pitch) at normal (average) speech to p1 and the pitch at newly sounding the same sound to p2. In general, the higher the p2, the higher the amplitude. Therefore, to normalize the amplitude to the amplitude at p1, p1 / p2 is calculated by adopting a new gradient control constant value. Way. The method of claim 1 or 2, wherein after the slope of the amplitude is determined, the feature vectors are HMM decoded.