METHOD FOR EVALUATING THE SPEECH QUALITY OF A SPEECH SIGNAL USING A HEARING DEVICE

Description

[0001] The invention relates to a method for evaluating the speech quality of a speech signal by means of a hearing device, wherein a sound containing the speech signal is recorded from an environment of the hearing device by means of an acousto-electrical input transducer of the hearing device and converted into an input audio signal, wherein at least one property of the speech signal is quantitatively recorded by analyzing the input audio signal by means of signal processing.

[0002] An important task in the use of hearing devices, such as hearing aids, headsets or communication devices, is often to output a speech signal to the user of the hearing device as precisely as possible, i.e. in particular in a way that is as acoustically intelligible as possible. To this end, background noise from an audio signal generated using a sound with a speech signal is often suppressed in order to emphasize the signal components that represent the speech signal and thus improve its intelligibility. However, noise suppression algorithms can often reduce the sound quality of a resulting output signal. Signal processing of the audio signal can in particular result in artifacts and / or the auditory impression is generally perceived as less natural.

[0003] Noise suppression is usually performed based on parameters that primarily affect the noise or the overall signal, such as a signal-to-noise ratio (SNR), a noise floor (noise floor), or even an audio signal level. However, this approach to controlling noise suppression can ultimately result in noise suppression being applied even when, despite noticeable background noise, it would not be necessary because some speech remains clearly intelligible despite the background noise. In this case, the risk of a deterioration in sound quality, e.g., due to noise suppression artifacts, is taken unnecessarily. Conversely, a speech signal that is only slightly overlaid with noise and, provided the associated audio signal has a good SNR, may also exhibit poor speech quality if the speaker's articulation is weak.

[0004] This could be avoided if algorithms for noise suppression in particular, but also signal processing in general, were controlled in a hearing aid depending on the quality of a speech signal component in the audio signal to be processed. However, this requires that such quality be measurable and detectable.

[0005] US 2004 / 0 167 774 A1 describes a method for analyzing and evaluating voices. A test speech signal is processed using a hearing model, at least one speech quality feature is determined from the test speech signal, and this speech quality feature is compared with a corresponding base speech quality feature. Based on this, a measure of the speech quality of the test speech signal can be determined.

[0006] In A. H. Andersen et al., "Nonintrusive Speech Intelligibility Prediction Using Convolutional Neural Networks," IEEE / ACM Transactions on Audio, Speech and Language Processing, IEEE, USA, Vol. 26, No. 10, October 1, 2018, pages 1925–1939, a convolutional neural network (CNN) is used for non-intrusive speech intelligibility prediction (SIP), which does not require a clean (i.e., noise-free) speech signal for predictions. The CNN architecture used is similar to existing SIP algorithms, making the trained CNN weights easy and meaningful to interpret. The proposed method shows high prediction performance compared to existing intrusive and non-intrusive SIP algorithms. This demonstrates the potential of deep learning for speech intelligibility prediction.

[0007] US 2018 / 0 255 406 A1 discloses a hearing aid comprising a plurality of microphones, including a first microphone for generating a first microphone input signal, and a processor for processing input signals into an electrical output signal. The hearing aid further comprises a receiver for converting the electrical output signal into an output sound signal, and a control unit operatively connected to the plurality of microphones, wherein the control unit estimates a speech intelligibility indicator based on one or more microphone input signals. The control unit is further configured to control the processor based on the speech intelligibility indicator. To estimate said indicator, a pitch parameter of a first audio source is estimated. The speech intelligibility indicator is based on the pitch parameter and a direction of the first audio source.

[0008] US Pat. No. 7,165,025 describes an articulation analysis for use in speech quality assessment. Articulation analysis is based on a comparison between the performance of articulatory and non-articulatory frequency ranges of a speech signal, with speech quality being assessed based on said comparison. Neither the source speech signal nor an estimate of it is used in the articulation analysis.

[0009] The invention is therefore based on the object of providing a method by which the quality of a speech component in an audio signal to be processed by a hearing device can be objectively assessed. The invention is further based on the object of providing a hearing device configured to objectively assess the quality of a speech component contained in an internal audio signal.

[0010] The first-mentioned object is achieved according to the invention by a method for evaluating the speech quality of a speech signal by means of a hearing device, wherein a sound containing the speech signal is recorded from an environment of the hearing device by means of an acousto-electrical input transducer of the hearing device and converted into an input audio signal, wherein at least one articulatory property of the speech signal is quantitatively recorded by analyzing the input audio signal by means of signal processing, in particular signal processing of the hearing device and / or an auxiliary device connectable to the hearing device, and wherein a quantitative measure of the speech quality is derived as a function of the at least one articulatory property. Advantageous and partly inventive embodiments are the subject of the dependent claims and the following description.

[0011] The second object is achieved according to the invention by a hearing device which comprises an acousto-electrical input transducer and a signal processing device, in particular having a signal processor, wherein the acousto-electrical input transducer is configured to record a sound from an environment of the hearing device and to convert it into an input audio signal, and wherein the signal processing device is configured to quantitatively detect at least one articulatory property of a portion of a speech signal contained in the input audio signal by analyzing the input audio signal and to derive a quantitative measure of the speech quality as a function of the at least one articulatory property according to the method described above.

[0012] The hearing device according to the invention shares the advantages of the method according to the invention, which can be carried out in particular using the hearing device according to the invention. The advantages mentioned below for the method and its further developments can be applied analogously to the hearing device.

[0013] An acoustoelectrical input transducer includes, in particular, any transducer configured to generate an electrical audio signal from ambient sound, such that air movements and air pressure fluctuations caused by the sound at the location of the transducer are reproduced by corresponding oscillations of an electrical quantity, in particular a voltage, in the generated audio signal. In particular, the acoustoelectrical input transducer can be a microphone.

[0014] The signal processing is carried out in particular by means of a corresponding signal processing device, which is configured by means of at least one signal processor to carry out the calculations and / or algorithms provided for signal processing. The signal processing device is arranged in particular on the hearing device. However, the signal processing device can also be arranged on an auxiliary device which is configured for connection to the hearing device for data exchange, e.g. a smartphone, a smartwatch, or similar. The hearing device can then, for example, transmit the input audio signal to the auxiliary device, and the analysis is carried out using the computing resources provided by the auxiliary device. Finally, the quantitative measure can be transmitted back to the hearing device as the result of the analysis.

[0015] The analysis can be performed directly on the input audio signal or based on a signal derived from the input audio signal. Such a signal can be provided, in particular, by the isolated speech signal component, but also by an audio signal, such as that generated in a hearing aid through a feedback loop using a compensation signal to compensate for acoustic feedback, or similar, or by a directional signal generated based on another input audio signal from another input transducer.

[0016] An articulatory property of a speech signal includes the precision of formants, especially vowels, and the dominance of consonants, especially fricatives and / or plosives. It can be stated that the higher the precision of the formants or the higher the dominance and / or precision of consonants, the higher the speech quality. A prosodic property of a speech signal includes, in particular, the temporal stability of a fundamental frequency of the speech signal and the relative sound intensity of accents.

[0017] Sound generation typically involves three physical components of a sound source: a mechanical oscillator, such as a string or membrane, which causes the air surrounding the oscillator to vibrate; an excitation of the oscillator (e.g., by plucking or bowing); and a resonating body. The excitation causes the oscillator to oscillate, causing the air surrounding the oscillator to vibrate under pressure, propagating as sound waves. In this case, the mechanical oscillator usually excites not only oscillations of a single frequency, but oscillations of various frequencies, with the spectral composition of the propagating oscillations determining the sound. The frequencies of certain oscillations are often given as integer multiples of a fundamental frequency and are referred to as "harmonics" or overtones of this fundamental frequency.However, more complex spectral patterns can also develop, meaning that not all generated frequencies can be represented as harmonics of the same fundamental frequency. The resonance of the generated frequencies in the resonant chamber is also relevant for the sound image, since certain frequencies generated by the oscillator are often attenuated in the resonant chamber relative to the dominant frequencies of a sound.

[0018] Applied to the human voice, this means that the mechanical oscillator is provided by the vocal cords, and their excitation in the air flowing past the vocal cords from the lungs, with the resonance chamber being formed primarily by the throat and mouth. The fundamental frequency of a male voice is usually in the range of 60 Hz to 150 Hz, for women it is usually in the range of 150 Hz to 300 Hz. As a result of the anatomical differences between individuals, both in terms of their vocal cords and particularly in terms of the throat and mouth, different sounding voices initially develop. By changing the volume and geometry of the mouth through appropriate jaw and lip movements, the resonance chamber can be altered in such a way that frequencies characteristic of the production of vowels are formed, so-called formants.These lie within fixed frequency ranges (the so-called "formant ranges") for individual vowels, with a vowel usually already clearly audibly distinguished from other sounds by the first two formants F1 and F2 of a series of often four formants (cf. "vowel triangle" and "vowel trapezoid"). The formants develop independently of the fundamental frequency, i.e., the frequency of the fundamental vibration.

[0019] In this sense, the precision of formants is to be understood as a degree of concentration of the acoustic energy on mutually definable formant ranges, in particular on individual frequencies in the formant ranges, and a resulting determinability of the individual vowels on the basis of the formants.

[0020] To produce consonants, the airflow passing the vocal cords is partially or completely blocked at at least one point, which, among other things, creates turbulence in the airflow. This is why only some consonants can be assigned a formant structure as clear as vowels, while others exhibit a more broadband frequency structure. However, consonants can also be assigned specific frequency bands in which the acoustic energy is concentrated. Due to the rather percussive "noise-like" nature of consonants, these frequency bands generally lie above the formant ranges of vowels, namely primarily in the range from approximately 2 to 8 kHz, while the ranges of the most important formants F1 and F2 of vowels generally end at approximately 1.5 kHz (F1) and 4 kHz (F2), respectively.The precision of consonants is determined in particular by the degree of concentration of acoustic energy on the corresponding frequency ranges and the resulting determinability of the individual consonants.

[0021] However, the distinguishability of the individual components of a speech signal, and thus the ability to resolve these components, depends not only on articulatory aspects. While these primarily concern the acoustic precision of the smallest isolated sound events of speech, the so-called phonemes, prosodic aspects also determine speech quality, since intonation and accentuation, particularly across multiple segments—i.e., multiple phonemes or phoneme groups—can impart a particular meaning to a statement. For example, by raising the pitch at the end of a sentence to clarify a question, or by stressing a specific syllable in a word to distinguish different meanings (cf. "umfahren" vs. "umfahren"), or by stressing a word to emphasize its prominence.In this respect, speech quality for a speech signal can also be quantitatively assessed using prosodic properties, in particular those just mentioned, by determining, for example, measures for a temporal variation of the pitch of the voice, i.e. its fundamental frequency, and for the clarity of a contrast of the amplitude and / or level maxima.

[0022] Based on one or more of the above-mentioned and / or other quantitatively recorded articulatory and / or prosodic properties of the speech signal, the quantitative measure of speech quality can be derived.

[0023] According to the invention, a characteristic correlated with the precision of given formants of vowels in the speech signal, a characteristic correlated with the dominance of consonants, in particular fricatives, in the speech signal, and / or a characteristic correlated with the precision of the transitions between voiced and unvoiced sounds is recorded as an articulatory property of the speech signal. The quantitative measure of speech quality can then be provided directly by the said recorded characteristic or can be formed on the basis of it, e.g., by weighting two characteristics for different formants or similar, or by weighting, i.e., by calculating a weighted average, at least two different of the said characteristics relative to each other. The quantitative measure of speech quality thus refers to the speech production of a speaker who, from a pronunciation perceived as "clean," has deficits (such as, for example,Lisping or mumbling) to speech defects, which reduce the quality of speech accordingly.

[0024] In contrast to quantities that are related to the propagation of speech in an environment, such as the Speech Intellegibility Index (SII), which weights the individual speech and noise components band by band, or the Speech Transmission Index (STI), which uses a test signal simulating the modulations of human speech to record the effect of a transmission channel on the modulation depth, the present measure for the is in particular independent of the external properties of a transmission channel, such as propagation in a possibly reverberating room or a noisy environment, but preferably only dependent on the intrinsic properties of speech production by the speaker.

[0025] This means in particular that in quiet environments and / or environments with only a low background noise, a reduced speech quality (relative to a reference value which is preferably set for speech quality perceived as "very good") is detected.

[0026] Within the scope of the invention, in order to detect the characteristic value correlated with the dominance of consonants in the speech signal, a first energy contained in a low frequency range is calculated, a second energy contained in a higher frequency range lying above the low frequency range is calculated, and the correlated characteristic value is formed on the basis of a ratio and / or a ratio of the first energy and the second energy weighted over the respective bandwidths of the said frequency ranges. In particular, a temporal smoothing of the speech signal can take place beforehand. To calculate the first and second energy, the input audio signal can in particular be divided into the low and the higher frequency range, e.g. by means of a filter bank and optionally by means of a corresponding selection of individual resulting frequency bands.Here, the lower frequency range is selected such that it lies within the frequency interval [0 Hz, 2.5 kHz], particularly preferably within the frequency interval [0 Hz, 2 kHz]. Furthermore, the higher frequency range is selected such that it lies within the frequency interval [3 kHz, 10 kHz], particularly preferably within the frequency interval [4 Hz, 8 kHz].

[0027] Within the scope of the invention, in order to detect the characteristic value correlated with the precision of the transitions between voiced and unvoiced sounds, a differentiation between voiced and unvoiced time sequences is carried out on the basis of a correlation measurement and / or on the basis of a zero-crossing rate of the input audio signal or a signal derived from the input audio signal, wherein a transition from a voiced time sequence to an unvoiced time sequence or from an unvoiced time sequence to a voiced time sequence is determined, the energy contained in the voiced or unvoiced time sequence before the transition is determined for at least one frequency range, and the energy contained in the unvoiced or voiced time sequence after the transition is determined for the at least one frequency range, and the characteristic value is determined on the basis of the energy before the transition and on the basis of the energy after the transition.

[0028] This means in particular: First, the voiced and unvoiced time sequences of the speech signal in the input audio signal are determined, and from this a transition from voiced to unvoiced or from unvoiced to voiced is identified. For at least one frequency range, which is specified in particular on the basis of empirical findings for the precision of the transitions, the energy before the transition in the frequency domain is then determined for the input audio signal or for a signal derived from it. This energy can, for example, be taken from the voiced or unvoiced time sequence immediately before the transition. Likewise, the energy in the relevant frequency range after the transition is determined, for example, from the unvoiced or voiced time sequence following the transition.

[0029] Based on these two energies, a characteristic value can now be determined which, in particular, allows a statement to be made about a change in the energy distribution at the transition. This characteristic value can, for example, be determined as a quotient or a relative deviation of the two energies before and after the transition. However, the characteristic value can also be calculated as a comparison of the energy before and after the transition with the total (broadband) signal energy. In particular, the energies before and after the transition can also be determined for a further frequency range, so that the characteristic value can also be determined based on the energies before and after the transition in the further frequency band, e.g. as a rate of change of the energy distribution across the relevant frequency ranges across the transition (i.e. a comparison of the distribution of energies in both frequency ranges before the transition with the distribution after the transition).

[0030] Using said characteristic value, the characteristic value for the measure of speech quality that correlates with the precision of the transitions can then be determined. For this purpose, the characteristic value can be used directly, or the characteristic value can be compared with a reference value previously determined for good articulation, particularly on the basis of corresponding empirical knowledge (e.g., as a quotient or relative deviation). The concrete design, particularly with regard to the frequency ranges and limit or reference values ​​to be used, can generally be based on empirical results regarding the corresponding significance of the respective frequency bands or groups of frequency bands. In particular, frequency bands 13 to 24, preferably 16 to 23 of the Bark scale can be used as the at least one frequency range. A frequency range of lower frequencies can be used as a further frequency range.

[0031] Within the scope of the invention, to detect the characteristic value correlated with the precision of predetermined formants of vowels in the speech signal, the acoustic energies of the speech signal concentrated in at least two different formant ranges (or quantities correlated with said energies) are compared with one another. Particularly preferably, a signal component of the speech signal is determined in at least one formant range in the frequency domain, a signal quantity correlated with the level is determined for the signal component of the speech signal in at least one formant range, and the characteristic value is determined based on a maximum value and / or based on a temporal stability of the signal quantity correlated with the level.In particular, the frequency range of the first formant F1 (preferably 250 Hz to 1 kHz, particularly preferably 300 Hz to 750 Hz) or the second formant F2 (preferably 500 Hz to 3.5 kHz, particularly preferably 600 Hz to 2.5 kHz) can be selected as the at least formant range, or two formant ranges of the first and second formants can be selected. In particular, several first and / or second formant ranges assigned to different vowels (i.e., the frequency ranges assigned to the first or second formant of the respective vowel) can also be selected. The signal component is then determined for the selected formant range(s), and a signal magnitude of the respective signal component correlated with the level is determined. The signal magnitude can be given by the level itself or by the maximum signal amplitude, which may be suitably smoothed.Based on a temporal stability of the signal size, which in turn can be determined by a variance of the signal size over a suitable time window, and / or based on a deviation of the signal size from its maximum value over a suitable time window, a statement can now be made about the precision of formants in that a low variance and a low deviation from the maximum level for an articulated sound (the length of the time window can be chosen depending on the length of an articulated sound) indicate a high precision.

[0032] Advantageously, by analyzing the input audio signal by means of signal processing, at least one prosodic property of the speech signal is further quantitatively recorded, and the quantitative measure for the speech quality is additionally determined as a function of the at least one prosodic property of the speech signal (18). Preferably, the fundamental frequency of the speech signal is recorded in a time-resolved manner, and a parameter characteristic of the temporal stability of the fundamental frequency is determined as a prosodic property of the speech signal. This parameter can be determined, for example, based on a relative deviation of the fundamental frequency accumulated over time, or by recording a number of maxima and minima of the fundamental frequency over a predetermined period of time. The temporal stability of the fundamental frequency is vais important for the monotony of speech melody and accentuation, which is why a quantitative recording also allows a statement about the speech quality of the speech signal.

[0033] Preferably, a quantity correlated with the volume, in particular an amplitude and / or a level, is recorded for the speech signal in a time-resolved manner, in particular by a corresponding analysis of the input audio signal or a signal derived therefrom, wherein a quotient of a maximum value of the quantity correlated with the volume and an average value of the said quantity determined over the predetermined period is formed, and wherein a characteristic quantity is determined as a prosodic property of the speech signal as a function of said quotient, which is formed from the maximum value and the average value of the quantity correlated with the volume over the predetermined period. In this way, a statement about a definition of the accentuation can be made on the basis of the indirectly recorded volume dynamics of the speech signal.

[0034] In an advantageous embodiment, at least two parameters, each characteristic of articulatory and / or prosodic properties, are determined based on the analysis of the input audio signal, wherein the quantitative measure of speech quality is formed based on a product of these parameters and / or based on a weighted mean and / or a maximum or minimum value of these parameters. This is particularly advantageous when a single measure of speech quality is required or desired, or when a single measure that is intended to capture all articulatory or all prosodic properties is desired.

[0035] Preferably, before detecting the at least one articulatory and / or prosodic property of the speech signal, speech activity is detected and / or an SNR is determined in the input audio signal, wherein an analysis is carried out with regard to the at least one articulatory and / or prosodic property of the speech signal as a function of the detected speech activity or the determined SNR. In this way, the analysis of the speech quality of the speech signal can be restricted to those cases in which a speech signal is actually present or in which the SNR is in particular above a predetermined limit value, so that it can be assumed that sufficiently good recognition of the signal components of the speech signal in the input audio signal is even possible in the first place in order to carry out a corresponding evaluation. Conversely, with conventional signal processing, no emphasis or similar measures are usually taken to achieve a sufficiently high SNR.of a speech signal, although poor speech quality, i.e. weak articulation and / or low expression of prosodic features such as accentuation, would benefit from improvement by means of signal processing.

[0036] The hearing device is preferably designed as a hearing aid. The hearing aid can be a monaural device or a binaural device with two local devices, which the user of the hearing aid wears on their right and left ears, respectively. In particular, in addition to the aforementioned input transducer, the hearing aid can also have at least one further acousto-electrical input transducer, which converts the ambient sound into a corresponding further input audio signal, so that the quantitative detection of the at least one articulatory and / or prosodic property of a speech signal can be achieved by analyzing a plurality of input audio signals involved. In the case of a binaural device, two of the input audio signals used can each be generated in different local units of the hearing aid (i.e., in the left and right ears, respectively).The signal processing device can in particular comprise signal processors of both local units, wherein preferably locally generated measures for the speech quality are standardized in a suitable manner by averaging or a maximum or minimum value for both local units, depending on the articulatory and / or prosodic property under consideration.

[0037] An exemplary embodiment of the invention is explained in more detail below with reference to a drawing. The drawings schematically show: Fig. 1 shows a circuit diagram of a hearing aid which detects a sound with a speech signal, and Fig. 2 shows a block diagram of a method for determining a quantitative measure of the speech quality of the speech signal according to Fig. 1 .

[0038] Corresponding parts and sizes are provided with the same reference numerals in all figures.

[0039] In Figure 1A schematic circuit diagram of a hearing device 1 is shown, which in this case is designed as a hearing aid 2. The hearing aid 2 has an acousto-electrical input transducer 4, which is configured to convert a sound 6 from the environment of the hearing aid 2 into an input audio signal 8. A configuration of the hearing aid 2 with a further input transducer (not shown), which generates a corresponding further input audio signal from the sound 6 from the environment, is also conceivable. The hearing aid 2 is designed as a stand-alone, monaural device. Equally conceivable is a configuration of the hearing aid 2 as a binaural hearing aid with two local devices (not shown), which are to be worn by the user of the hearing aid 2 on their right and left ears, respectively.

[0040] The input audio signal 8 is fed to a signal processing device 10 of the hearing aid 2, in which the input audio signal 8 is processed in particular according to the audiological requirements of the user of the hearing aid 2 and is thereby amplified and / or compressed, for example, frequency band by frequency. The signal processing device 10 is designed for this purpose in particular by means of a corresponding signal processor (in Figure 1 (not shown in detail) and a working memory addressable via the signal processor. Any preprocessing of the input audio signal 8, such as an A / D conversion and / or preamplification of the generated input audio signal 8, shall be considered part of the input converter 4.

[0041] By processing the input audio signal 8, the signal processing device 10 generates an output audio signal 12, which is converted into an output sound signal 16 of the hearing aid 2 by an electroacoustic output transducer 14. The input transducer 4 is preferably a microphone, while the output transducer 14 is, for example, a loudspeaker (such as a balanced metal case receiver), but can also be a bone conduction receiver or similar.

[0042] The sound 6 of the environment of the hearing aid 2, which is detected by the input transducer 4, includes, among other things, a speech signal 18 of a speaker not shown in detail, as well as further sound components 20, which can in particular comprise directed and / or diffuse interference noises (interference noise or background noise), but can also include noises which, depending on the situation, could be regarded as a useful signal, for example music or acoustic warning or information signals relating to the environment.

[0043] The signal processing of the input audio signal 8 performed in the signal processing device 10 to generate the output audio signal 12 can, in particular, comprise suppressing the signal components that suppress the noise contained in the sound 6, or a relative amplification of the signal components representing the speech signal 18 compared to the signal components representing the other sound components 20. In particular, frequency-dependent or broadband dynamic compression and / or amplification as well as noise suppression algorithms can also be applied.

[0044] In order to make the signal components in the input audio signal 8, which represent the speech signal 18, as audible as possible in the output audio signal 12, and still be able to convey to the user of the hearing aid 2 in the output sound 16 the most natural auditory impression possible, a quantitative measure for the speech quality of the speech signal 18 is to be determined in the signal processing device 10 for controlling the algorithms to be applied to the input audio signal 8. This is to be done using Figure 2 described.

[0045] Figure 2 shows in a block diagram a processing of the input audio signal 8 of the hearing aid 2 according to Figure 2First, a speech activity detection (VAD) is performed for the input audio signal 8. If no significant speech activity is present (path "n"), the signal processing of the input audio signal 8 to generate the output audio signal 12 is carried out using a first algorithm 25. The first algorithm 25 evaluates signal parameters of the input audio signal 8, such as level, background noise, transients, or the like, in a predetermined manner, broadband and / or in particular frequency band-wise, and from this determines individual parameters, e.g., frequency band-wise gain factors and / or compression characteristics (i.e., in particular, knee point, ratio, attack, release), which are to be applied to the input audio signal 8.

[0046] In particular, the first algorithm 25 can also provide a classification of a hearing situation realized in the sound 6 and, depending on the classification, adjust individual parameters, possibly as a hearing program intended for a specific hearing situation. Furthermore, the individual audiological requirements of the user of the hearing aid 2 can also be taken into account for the first algorithm 25 in order to compensate for the user's hearing impairment as effectively as possible by applying the first algorithm 25 to the input audio signal 8.

[0047] However, if significant speech activity is detected during the detection of speech activity VAD (path "y"), an SNR is next determined and compared with a predetermined limit value TH SNR. If the SNR is not above the limit value, i.e. SNR ≤ TH SNR , the first algorithm 25 is again applied to the input audio signal 8 to generate the output audio signal 12. However, if the SNR is above the predetermined limit value TH SNR , i.e. SNR > TH SNR , a quantitative measure 30 for the speech quality of the speech component 18 contained in the input audio signal 8 is determined for further processing of the input audio signal 8 in the manner described below. For this purpose, articulatory and / or prosodic properties of the speech signal 18 are quantitatively recorded.The term speech signal component 26 contained in the input audio signal 8 refers to those signal components of the input audio signal 8 which represent the speech component 18 of the sound 6 from which the input audio signal 8 is generated by means of the input transducer 4.

[0048] To determine the said quantitative measure 30, the input audio signal 8 is divided into individual signal paths.

[0049] For a first signal path 32 of the input audio signal 8, a center wavelength λc is first determined and compared with a predetermined limit value for the center wavelength Th λ. If, based on the said limit value for the center wavelength Th λ, it is determined that the signal components in the input audio signal 8 are sufficiently high-frequency, the signal components are selected in the first signal path 32, possibly after a suitably selected temporal smoothing (not shown), for a low frequency range LF and a higher frequency range HF lying above the low frequency range LF. A possible division can, for example, be such that the low frequency range LF includes all frequencies f N ≤ 2500 Hz, in particular f N ≤ 2000 Hz, and the higher frequency range HF includes frequencies f H with 2500 Hz < f H ≤ 10000 Hz, in particular 4000 Hz ≤ f H ≤ 8000 Hz or 2500 Hz < f H ≤ 5000 Hz.

[0050] The selection can be carried out directly in the input audio signal 8, or can also be carried out in such a way that the input audio signal 8 is divided into individual frequency bands by means of a filter bank (not shown), whereby individual frequency bands are assigned to the low or higher frequency range NF or HF depending on the respective band limits.

[0051] Subsequently, a first energy E1 is determined for the signal contained in the low frequency range LF and a second energy E2 is determined for the signal contained in the higher frequency range HF. A quotient QE is then formed from the second energy as the numerator and the first energy E1 as the denominator. With a suitably selected lower and higher frequency range LF, HF, the quotient QE can now be used as a parameter 33 which is correlated with the dominance of consonants in the speech signal 18. The parameter 33 thus enables a statement to be made about an articulatory property of the speech signal components 26 in the input audio signal 8. For example, for a value of the quotient QE >> 1 (i.e. QE > Th QE with a predetermined, not shown limit value Th QE >> 1) a high dominance for consonants can be inferred, whereas for a value QE < 1 a low dominance can be inferred.

[0052] In a second signal path 34, a distinction 36 is made in the input audio signal 8 between voiced time sequences V and unvoiced time sequences UV based on correlation measurements and / or a zero-crossing rate of the input audio signal 8. Based on the voiced and unvoiced time sequences V and UV, respectively, a transition TS from a voiced time sequence V to an unvoiced time sequence UV is determined. The length of a voiced or unvoiced time sequence can be, for example, between 10 and 80 ms, in particular between 20 and 50 ms.

[0053] For at least one frequency range (e.g. a selection of particularly meaningful frequency bands determined to be suitable, e.g. frequency bands 16 to 23 of the Bark scale, or frequency bands 1 to 15 of the Bark scale), an energy Ev is determined for the voiced time sequence V before the transition TS and an energy En for the unvoiced time sequence UV after the transition TS. In particular, corresponding energies before and after the transition TS can also be determined separately for more than one frequency range. It is now determined how the energy changes at the transition TS, e.g. by a relative change ΔE TS or by a quotient (not shown) of the energies Ev, En before and after the transition TS.

[0054] The measure of the change in energy, i.e., in this case, the relative change, is then compared with a limit value Th E for energy distribution at transitions, previously determined for good articulation. In particular, a characteristic 35 can be calculated based on a ratio of the relative change ΔE TS and the said limit value Th E or based on a relative deviation of the relative change ΔE TS from this limit value Th E. Said characteristic 35 is correlated with the articulation of the transitions between voiced and unvoiced sounds in the speech signal 18 and thus provides information about a further articulatory property of the speech signal components 26 in the input audio signal 8.In general, the statement applies that the transition between voiced and unvoiced time sequences is articulated more precisely the faster, i.e. the more temporally definable, a change in energy distribution occurs across the frequency ranges relevant for voiced and unvoiced sounds.

[0055] For parameter 35, however, an energy distribution in two frequency ranges (e.g., the above-mentioned frequency ranges according to the Bark scale, or also in the lower and higher frequency ranges LF, HF) can also be considered, e.g., using a quotient of the respective energies or a comparable parameter, and a change in the quotient or parameter across the transition can be used for the parameter. For example, a rate of change of the quotient or parameter can be determined and compared with a previously determined suitable reference value for the rate of change.

[0056] To generate parameter 35, transitions of unvoiced time sequences can also be considered in an analogous manner. The specific design, particularly with regard to the frequency ranges and limit or reference values ​​to be used, can generally be determined based on empirical results regarding the corresponding significance of the respective frequency bands or groups of frequency bands.

[0057] In a third signal path 38, a fundamental frequency f G of the speech signal component 26 is detected in the input audio signal 8 in a time-resolved manner, and a time stability 40 is determined for said fundamental frequency f G based on a variance of the fundamental frequency f G. The time stability 40 can be used as a parameter 41, which enables a statement to be made about a prosodic property of the speech signal components 26 in the input audio signal 8. A greater variance in the fundamental frequency f G can be used as an indicator of better speech intelligibility, while a monotonous fundamental frequency f G has a lower speech intelligibility.

[0058] In a fourth signal path 42, a level LVL is recorded in a time-resolved manner for the input audio signal 8 and / or for the speech signal component 26 contained therein, and a temporal average MN LVL is calculated over a period 44 predetermined in particular on the basis of corresponding empirical findings. Furthermore, the maximum MX LVL of the level LVL is determined over the period 44. The maximum MX LVL of the level LVL is then divided by the temporal average MN LVL of the level LVL, thus determining a characteristic variable 45 correlated with a volume of the speech signal 18, which allows further information about a prosodic property of the speech signal components 26 in the input audio signal 8. Instead of the level LVL, another variable correlated with the volume and / or energy content of the speech signal component 26 can also be used.

[0059] The parameters 33, 35, 41, and 45, respectively, determined in the first to fourth signal paths 32, 34, 38, 42 as described above, can now each be used individually as the quantitative measure 30 for the quality of the speech component 18 contained in the input audio signal 8. Depending on the quality of the speech component 18 contained in the input audio signal 8, a second algorithm 46 is then applied to the input audio signal 8 for signal processing. The second algorithm 46 can be derived from the first algorithm 25 by a corresponding change in one or more signal processing parameters depending on the respective quantitative measure 30, or it can provide a completely independent hearing program.

[0060] In particular, a single value can be determined as a quantitative measure 30 for the speech quality based on the parameters 33, 35, 41 or 45 determined as described, e.g. by a weighted average or a product of the parameters 33, 35, 41, 45 (in Fig. 2 (schematically represented by combining parameters 33, 35, 41, 45). The weighting of the individual parameters can be carried out in particular using previously empirically determined weighting factors, which can be determined based on the significance of the articulatory or prosodic property captured by the respective parameter for speech quality.

[0061] Although the invention has been illustrated and described in detail by the preferred embodiment, the invention is not limited to the disclosed examples and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention. List of reference symbols

[0062] 1 Hearing device 2 Hearing aid 4 Input transducer 6 Ambient sound 8 Input audio signal 10 Signal processing device 12 Output audio signal 14 Output transducer 16 Output sound 18 Speech signal 20 Sound components 25 First algorithm 26 Speech signal component 30 Quantitative measure of speech quality 32 First signal path 33 Parameter 34 Second signal path 35 Parameter 36 Differentiation 38 Third signal path 40 Time stability 41 Parameter 42 Fourth signal path 44 Time period 45 Parameter 46 Second algorithm ΔE TS relative change (of energy at the transition) λ C center wavelength E1first energy E2second energy EvEnergy (before the transition) EnEnergy (after the transition) f G Fundamental frequency LVL Level HF Higher frequency range MN LVL Time average (of the level) MX LVL Maximum of the level NF Lower frequency range QE Quotient SNR Signal-to-noise ratio (SNR) Th λ Limit (for the centroid wavelength) Th E Limit (for relative change in energy) TH SNR Limit (for the SNR) TS Transition V Voiced time sequence VA Detection of speech activity UV Unvoiced time sequence

Claims

1. Method for rating the speech quality of a speech signal (18) by way of a hearing device (1), - wherein an acousto-electric input transducer (4) of the hearing device (1) records a sound (6) containing the speech signal (18) from surroundings of the hearing device (1) and converts it into an input audio signal (8), - wherein at least one articulatory property of the speech signal (18) is quantitatively acquired through analysis of the input audio signal (8) by way of a signal processing operation, and - wherein a quantitative measure (30) of the speech quality is derived on the basis of the at least one articulatory property wherein - a characteristic variable correlated with the precision of predefined formants of vowels in the speech signal (18), and / or - a characteristic variable (31) correlated with the dominance of consonants, in particular fricatives, in the speech signal (18), and / or - a characteristic variable (35) correlated with the precision of transitions from voiced and unvoiced sounds is acquired as articulatory property of the speech signal (18), characterized in that, in order to acquire the characteristic variable (33) correlated with the dominance of consonants in the speech signal (18), - a first energy (E1) contained in a low frequency range (NF) is calculated, wherein the low frequency range (NF) is selected within the frequency interval [0 Hz, 2.5 kHz], - a second energy (E2) contained in a frequency range (HF) higher than the low frequency range (E2) is calculated, wherein the higher frequency range (HF) is selected within the frequency interval [3 Hz, 10 kHz], - and the characteristic variable is formed based on a ratio (QE), and / or a ratio weighted over the respective bandwidths of said frequency ranges (NF, HF), of the first energy (E1) and the second energy (E2), or in that, in order to acquire the characteristic variable (35) correlated with precision of the transitions from voiced and unvoiced sounds, - a distinction (36) is made between voiced temporal sequences (V) and unvoiced temporal sequences (UV) based on a correlation measurement and / or based on a zero crossing rate, - a transition (TS) from a voiced temporal sequence (V) to an unvoiced temporal sequence (UV) or from an unvoiced temporal sequence (UV) to a voiced temporal sequence (V) is ascertained, - the energy (Ev) contained in the voiced or unvoiced temporal sequence (V, UV) prior to the transition (TS) is ascertained for at least one frequency range, and the energy (En) contained in the unvoiced or voiced temporal sequence (UV, V) following the transition (TS) is ascertained for the at least one frequency range, and - the characteristic variable (35) is ascertained based on the energy (Ev) prior to the transition (TS) and based on the energy (En) following the transition (TS), or in that, in order to acquire the characteristic variable correlated with the precision of predefined formants of vowels in the speech signal (18), - a signal component of the speech signal (18) in at least one formant range in the frequency space is ascertained, - a signal variable correlated with the level is ascertained for the signal component of the speech signal (18) in the at least one formant range, and - the characteristic variable is ascertained based on a maximum value and / or based on a temporal stability of the signal variable correlated with the level.

2. Method according to Claim 1, - wherein furthermore at least one prosodic property of the speech signal (18) is quantitatively acquired through analysis of the input audio signal (8) by way of the signal processing operation, and - wherein the quantitative measure (30) of the speech quality is additionally ascertained on the basis of the at least one prosodic property of the speech signal (18).

3. Method according to Claim 2, wherein the fundamental frequency (fG) of the speech signal (18) is acquired in a temporally resolved manner, and wherein a characteristic variable (41) characteristic of the temporal stability (40) of the fundamental frequency (fG) is ascertained as prosodic property of the speech signal (18).

4. Method according to Claim 2 or Claim 3, wherein a variable (LVL) correlated with the volume is acquired in a temporally resolved manner for the speech signal (18), wherein a quotient of a maximum value (MXLVL) of the variable (LVL) correlated with the volume to a mean (MNLVL) of said variable (LVL), ascertained over a predefined time interval (44), is formed over the predefined time interval (44), and wherein a characteristic variable (45) is ascertained as prosodic property of the speech signal (18) on the basis of said quotient that is formed from the maximum value (MXLVL) and the mean (MNLVL) of the variable (VL) correlated with the volume over the predefined time interval (44).

5. Method according to one of the preceding claims, wherein at least two characteristic variables (33, 35, 41, 45) each characteristic of articulatory and / or prosodic properties are ascertained based on the analysis of the input audio signal (18), and wherein the quantitative measure (30) of the speech quality is formed based on a product of these characteristic variables (33, 35, 41, 45) and / or based on a weighted mean of these characteristic variables (33, 35, 41, 45).

6. Method according to one of the preceding claims, wherein speech activity (VAD) is detected and / or a signal-to-noise ratio (SNR) in the input audio signal (18) is ascertained before the at least one articulatory and / or prosodic property of the speech signal is acquired, and wherein analysis is performed with regard to the at least one articulatory and / or prosodic property of the speech signal (18) on the basis of the detected voice activity (VAD) or the ascertained signal-to-noise ratio (SNR).

7. Hearing device (1) comprising: - an acousto-electric input transducer (4) that is designed to record a sound (6) from surroundings of the hearing device (1) and to convert it into an input audio signal (8), and - a signal processing apparatus (10) that is designed to quantitatively acquire at least one articulatory property of a component, contained in the input audio signal (8), of a speech signal (18) based on analysis of the input audio signal (8) and to derive a quantitative measure (30) of the speech quality on the basis of the at least one articulatory property according to the method according to any of the preceding claims.

8. Hearing device (1) according to Claim 7, designed as a hearing aid (2).