3.2.1.

Time analysis

The time analysis focused on studying two issues. The first one concerns speech volume measurements, whereas the second one concerns speech rate.

Speech signal volume is one of the basic features, which is possible to be subjectively assessed by a human. In general, the literature identifies the term volume with signal power or its intensity. Volume is affected by the used microphone and preamplifier gain. Due to the fact that all recordings were completed using the same equipment, it is possible to objectively compare the speech volume of the studied persons. The experiments involved studying the volume of the “a” sound uttered by a group of people with PD and the control group. The calculations utilized the formulas contained in [12]. The outcome was the volume waveform expressed in decibels, as shown in Fig. 1.

Figure 1.

An example time waveform of the “a” sound with prolonged phonation (5 seconds) (a) and a volume waveform of a recording in dB in people with PD (b) and the control group (c).

The obtained waveform clearly shows the difference between the volume of a sound spoken by sick persons and the ones from the control group. In general, the speech of people affected by Parkinson’s disease is characterized by mutation. The so-called speech volume reduction occurs, which is confirmed by the studies conducted by the authors. Most likely this is caused by weakening of the larynx and lung muscles, which are responsible for the speech signal formation process. At the same time, please note the volume level changing over time. The obtained waveforms clearly indicate that people with diagnosed PD, colloquially speaking, tire faster, which results in the volume of the spoken text decreasing over time. Differences in the volume level are quite significant.

Due to the same linguistic material, in the case of analysing the signal rate, the research involved studying the duration of a utterance of “sad” and “joyful” nature, the average number of “pa” syllable uttered within a 5-second period, as well as average time intervals between syllables. The parameters are defined below and the results shown in table 1:

Table 1.

Exemplary time analysis results in PD patients.

Such measurements enable evaluating the rate of speech in the form of the number of words spoken per minute, due to the analysis of a text of the same linguistic content or conducting such an analysis using a method of counting the number of spoken “pa” syllables relative to a 5-second segment of speech. The obtained results indicate a general trend of an extending speech duration in people with PD. Both the average speech time is longer, as well as the number of “pa” syllables per an analysed 5-second speech segment are lower. In the case of pilot studies, it was particularly evident in patients with advanced Parkinson’s disease, falling within the upper part of the UPDRS. As a result, the rate of speech in people with PD is lower than in people from the control group. According to the studies, the average speaking rate is approximately 2.5 word per second, which amounts to 125 to 150 words per minute, whereas it should be emphasized that it may vary depending on the language, place of residence and many other factors [17]. In the case of the analysed texts, we had to deal with the same number of words. For a text with an optimistic undertone, the number of words was 86, while in the case of a sad text, this number was 90. Therefore, in general, Bradylalia, which is slow speech, can be determined in the case of people with diagnosed Parkinson’s disease. A change in the tension of laryngeal muscles impacts sound intensity, and the duration of ligament vibrations affects the decision in terms of sound duration [3].

An interesting aspect of the studies, which was not analysed widely by the authors, however, which draws attention to the issue concerning people with PD is the stability of time intervals between successively uttered “pa” syllables. An example time waveform with marked time intervals between subsequent syllables is shown in Fig. 2. Preliminary pilot studies enable concluding the lack of stability in the intervals between subsequent syllables and, in general, a lower rate of uttering syllables within a specific time unit.

Figure 2.

An example time waveform with marked intervals between subsequent “pa” syllables in a patient diagnosed with Parkinson’s disease.

Time parameters are the parameters primarily utilized to assess prosodic vocal properties. In the case of acoustic speech signal analysis these are frequently also parameters, which are possible to be estimated by means of hearing and are an important element in evaluating a PD patient. The application of ready parameter extracting procedures using analysis of the time domain will facilitate patient’s history taking and may possibly support the potential development or regression of a disease.

3.2.2.

Frequency analysis

The frequency analysis involved evaluating the changes in the fundamental frequency of the laryngeal tone, depending on the nature of the text uttered by an examined person. The fundamental frequency of a sound is determined in the literature using numerous methods [19-21]. Algorithms for detecting F₀ can be divided into methods operating within the time domain, frequency domain and both domains simultaneously. The common feature of the algorithms is detecting fundamental frequency only in the so-called voiced portions of the speech signal, due to the fact that non-voiced signals are of noise nature. The analysed signal frames shall be selected in terms of their voicing for this purpose. The standard is to use the autocorrelation function for this purpose, determining a certain threshold above which the frames are deemed voiced [22, 23]. Furthermore, the autocorrelation function was used by the authors to determine the fundamental frequency, because the determination method utilizing this function is, according to the researchers throughout the world, encumbered with the lowest error [24].

If a frame is deemed voiced, the value of its corresponding F_ac fundamental frequency is determined according to the relationship:

where k_max is the index of the second maximum of the autocorrelation function r(k), and f_p – the signal sampling frequency.

Fundamental frequency stability was assessed in the course of a long-term analysis, using texts of both joyful, as well as sad nature. A short-term analysis using the “a” sound with prolonged phonation was also conducted. In each of the analysed cases, in order to perform a digital analysis of a signal, segmentation is required. A speech signal analysed as a whole cannot be deemed a stationary signal, since its properties change significantly over time. However, if the signal is to be analysed against short time intervals (in the order of several milliseconds), during which the properties change only slightly, it can be treated as a quasi-stationary signal. Each frame is analysed individually, and the overall analysis is always a resultant analysis of the individual speech signal frames.

Fundamental frequency was calculated for 20 ms long speech signal frames with 5 ms long overlays. An example graph presenting a changing fundamental frequency of the laryngeal tone for the short-term analysis regarding a healthy and sick person is shown in Fig. 3.

Figure 3.

Example waveform of fundamental frequency variations during prolonged phonation of the “a” sound (short-term analysis) in a sick (top) and healthy (bottom) person.

The waveform of the fundamental frequency variations during prolonged phonation of the “a” sound indicates that the patients in the studied group exhibited higher instability of the fundamental frequency. It demonstrates the irregular work of the vocal folds, and more precisely, irregular frequency of their opening and closing. As a result, impaired laryngeal operation is the case for people with Parkinson’s disease [21]. It is particularly visible in the second half of the analysed time period. People with PD have difficulty in maintaining the utterance of a sound with the same pitch.

When looking for speech signal features, which correlate with the issue of early PD diagnostics, attention should be primarily paid to the fundamental frequency changes and the correlated statistical parameters, and bit the very value of the fundamental frequency, which is rather an indicator of the studied person’s identity.

The fundamental frequency was also extracted in the case of uttering texts of joyful and sad nature. The long-term analysis involved experiments aimed at evaluating speech monotony. All of this because it is widely believed that people with Parkinson’s disease are characterized by speech with a weaker emotional tinge. Table 2 shows the range of fundamental frequency changes when texts of various emotional undertone are uttered by a healthy woman, and a woman with diagnosed PD (both of similar age). The analysis time was about 30 seconds.

Table 2.

The range of fundamental frequency changes when texts of various emotional undertone are uttered by a healthy woman, and a woman with diagnosed PD (both of similar age)

Fig. 4 shows also the changes of fundamental frequency when texts of sad undertone (duration: 15 seconds) are uttered by a healthy person and a person diagnosed with Parkinson’s disease.

Figure 4.

Graph of laryngeal tone fundamental frequency changes for a text of sad nature (long-term analysis) uttered by a sick person (top) and a healthy person (bottom).

Based on the conducted experiments, it can be easily concluded that the fundamental frequency changes over a broader range in a healthy person. This is evidenced both by the range of fundamental frequency changes, as well as the standard deviation calculated for this parameter. As a result, the speech of people with PD is characterized by higher monotony and lack of modulation [21]. Although the read-out text was marked emotionally, people with diagnosed PD did not explicitly emphasize that fact during the recordings, which could be felt directly and subjectively in the course of the study.

Another study within the frequency analysis was the evaluation of speech voicing in people with Parkinson’s disease. The voicing of the frames was assessed for this purpose. The aforementioned autocorrelation function was used as the selection criterion. For this purpose, the authors formulated the so-called voicing coefficient, which is defined as follows

The variability graph for this parameter, depending on the nature of the spoken text, is shown below. The values on the left are for the group with diagnosed PD. The values on the right are for the control group.

Figure 5.

Voicing coefficient variability graph during the long-term analysis of statements of varying emotional undertone.

The results of the conducted experiments indicate lower values of the voicing coefficients in people with diagnosed Parkinson’s disease. Due to the fact that the study involved the same linguistic text, it can be speculated that in this case, the difference in experiment results is caused primarily by weaker operation of the vocal folds, as well as by the stiffness of the oral cavity, pharynx and larynx muscles in those patients. Furthermore, it can be caused by the issue of generating voiced vowels like “b” and inaccurate articulation in patients with PD [3].

The frequency analysis also involved comparing the amplitude spectra for individual words spoken by the studied persons. Examples of amplitude spectra in a decibel scale are shown in the figure below.

Figure 6.

Amplitude spectrum graph when uttering the word “aeroplan” (Eng. aeroplane). Healthy person on the left, sick person on the right.

Within the speech signal spectrum analysis (the word “aeroplan”), we can observe reduced vocal range, especially in terms of high frequencies. Amplitude for a person with diagnosed Parkinson’s disease is lower.

3.2.3.

Time-frequency analysis

The short-time Fourier transform (STFT) was used within the time-frequency analysis. This method involved successive division into segments of a signal section intended for analysing, each of which is independently subject to spectral analysis. Just like in the traditional case, in order to remove signal changes (cuts) at range extremes, various time windows are applied relative to the aforementioned segments. Shifting a time window along the signal determines its spectral content inside the time interval, the length of which is defined by the window width [22].

The acoustic waveform analysis based on spectrograms is an analysis of the long-term type. It is usually applied for recorded words and sentences, although it is also possible to use it for individual sounds. The research included analysing the word “aeroplan” (Eng. aeroplane), and the results’ examples are shown in Fig. 7.

Figure 7.

Spectrogram for the word “aeroplan”. Healthy person on the left, sick person on the right.

The spectrogram analysis indicates that Parkinson’s disease impacts the entire frequency range increasing their occurrence intensity. The spectrogram contrast degree (or otherwise – its clarity) is a measure of speech signal disturbances. The differences in the spectrograms most likely result from the varying proper closure of the vocal folds. Immediately after the vocal folds close, the harmonic structure is the richest, and the amplitude of all frequency components the greatest. As a result, the operation of the vocal folds is all the more efficient, the better the spectrogram contrast [24].