Algorithms for separating the speech of interfering talkers: evaluations with voiced sentences, and normal-hearing and hearing-impaired listeners

Two signal-processing algorithms, derived from those described by Stubbs and Summerfield [R.J. Stubbs and Q. Summerfield, J. Acoust. Soc. Am. 84, 1236-1249 (1988)], were used to separate the voiced speech of two talkers speaking simultaneously, at similar intensities, in a single channel. Both algorithms use fundamental frequency (FO) as the basis for segregation. One attenuates the interfering voice by filtering the cepstrum of the signal. The other is a hybrid algorithm that combines cepstral filtering with the technique of harmonic selection [T.W. Parsons, J. Acoust. Soc. Am. 60, 911-918 (1976)]. The algorithms were evaluated and compared in perceptual experiments involving listeners with normal hearing and listeners with cochlear hearing impairments. In experiment 1 the processing was used to separate voiced sentences spoken on a monotone. Both algorithms gave significant increases in intelligibility to both groups of listeners. The improvements were equivalent to an increase of 3-4 dB in the effective signal-to-noise ratio (SNR). In experiment 2 the processing was used to separate voiced sentences spoken with time-varying intonation. For normal-hearing listeners, cepstral filtering gave a significant increase in intelligibility, while the hybrid algorithm gave an increase that was on the margins of significance (p = 0.06). The improvements were equivalent to an increase of 2-3 dB in the effective SNR. For impaired listeners, no intelligibility improvements were demonstrated with intoned sentences. The decrease in performance for intoned material is attributed to limitations of the algorithms when FO is nonstationary.     

Evaluation of two voice-separation algorithms using normal-hearing and hearing-impaired listeners

Two signal-processing algorithms, designed to separate the voiced speech of two talkers speaking simultaneously at similar intensities in a single channel, were compared and evaluated. Both algorithms exploit the harmonic structure of voiced speech and require a difference in fundamental frequency (F0) between the voices to operate successfully. One attenuates the interfering voice by filtering the cepstrum of the combined signal. The other uses the method of harmonic selection [T. W. Parsons, J. Acoust. Soc. Am. 60, 911-918 (1976)] to resynthesize the target voice from fragmentary spectral information. Two perceptual evaluations were carried out. One involved the separation of pairs of vowels synthesized on static F0's; the other involved the recovery of consonant-vowel (CV) words masked by a synthesized vowel. Normal-hearing listeners and four listeners with moderate-to-severe, bilateral, symmetrical, sensorineural hearing impairments were tested. All listeners showed increased accuracy of identification when the target voice was enhanced by processing. The vowel-identification data show that intelligibility enhancement is possible over a range of F0 separations between the target and interfering voice. The recovery of CV words demonstrates that the processing is valid not only for spectrally static vowels but also for less intense time-varying voiced consonants. The results for the impaired listeners suggest that the algorithms may be applicable as components of a noise-reduction system in future digital signal-processing hearing aids. The vowel-separation test, and subjective listening, suggest that harmonic selection, which is the more computationally expensive method, produces the more effective voice separation.     

The benefit of binaural hearing in a cocktail party: effect of location and type of interferer

The "cocktail party problem" was studied using virtual stimuli whose spatial locations were generated using anechoic head-related impulse responses from the AUDIS database [Blauert et al., J. Acoust. Soc. Am. 103, 3082 (1998)]. Speech reception thresholds (SRTs) were measured for Harvard IEEE sentences presented from the front in the presence of one, two, or three interfering sources. Four types of interferer were used: (1) other sentences spoken by the same talker, (2) time-reversed sentences of the same talker, (3) speech-spectrum shaped noise, and (4) speech-spectrum shaped noise, modulated by the temporal envelope of the sentences. Each interferer was matched to the spectrum of the target talker. Interferers were placed in several spatial configurations, either coincident with or separated from the target. Binaural advantage was derived by subtracting SRTs from listening with the "better monaural ear" from those for binaural listening. For a single interferer, there was a binaural advantage of 2-4 dB for all interferer types. For two or three interferers, the advantage was 2-4 dB for noise and speech-modulated noise, and 6-7 dB for speech and time-reversed speech. These data suggest that the benefit of binaural hearing for speech intelligibility is especially pronounced when there are multiple voiced interferers at different locations from the target, regardless of spatial configuration; measurements with fewer or with other types of interferers can underestimate this benefit.     

Gain-induced speech distortions and the absence of intelligibility benefit with existing noise-reduction algorithms

Most noise-reduction algorithms used in hearing aids apply a gain to the noisy envelopes to reduce noise interference. The present study assesses the impact of two types of speech distortion introduced by noise-suppressive gain functions: amplification distortion occurring when the amplitude of the target signal is over-estimated, and attenuation distortion occurring when the target amplitude is under-estimated. Sentences corrupted by steady noise and competing talker were processed through a noise-reduction algorithm and synthesized to contain either amplification distortion, attenuation distortion or both. The attenuation distortion was found to have a minimal effect on speech intelligibility. In fact, substantial improvements (>80 percentage points) in intelligibility, relative to noise-corrupted speech, were obtained when the processed sentences contained only attenuation distortion. When the amplification distortion was limited to be smaller than 6 dB, performance was nearly unaffected in the steady-noise conditions, but was severely degraded in the competing-talker conditions. Overall, the present data suggest that one reason that existing algorithms do not improve speech intelligibility is because they allow amplification distortions in excess of 6 dB. These distortions are shown in this study to be always associated with masker-dominated envelopes and should thus be eliminated.     

Effects of fluctuating noise and interfering speech on the speech-reception threshold for impaired and normal hearing

The speech-reception threshold (SRT) for sentences presented in a fluctuating interfering background sound of 80 dBA SPL is measured for 20 normal-hearing listeners and 20 listeners with sensorineural hearing impairment. The interfering sounds range from steady-state noise, via modulated noise, to a single competing voice. Two voices are used, one male and one female, and the spectrum of the masker is shaped according to these voices. For both voices, the SRT is measured as well in noise spectrally shaped according to the target voice as shaped according to the other voice. The results show that, for normal-hearing listeners, the SRT for sentences in modulated noise is 4-6 dB lower than for steady-state noise; for sentences masked by a competing voice, this difference is 6-8 dB. For listeners with moderate sensorineural hearing loss, elevated thresholds are obtained without an appreciable effect of masker fluctuations. The implications of these results for estimating a hearing handicap in everyday conditions are discussed. By using the articulation index (AI), it is shown that hearing-impaired individuals perform poorer than suggested by the loss of audibility for some parts of the speech signal. Finally, three mechanisms are discussed that contribute to the absence of unmasking by masker fluctuations in hearing-impaired listeners. The low sensation level at which the impaired listeners receive the masker seems a major determinant. The second and third factors are: reduced temporal resolution and a reduction in comodulation masking release, respectively.     

Cochlear implant speech recognition with speech maskers

Speech recognition performance was measured in normal-hearing and cochlear-implant listeners with maskers consisting of either steady-state speech-spectrum-shaped noise or a competing sentence. Target sentences from a male talker were presented in the presence of one of three competing talkers (same male, different male, or female) or speech-spectrum-shaped noise generated from this talker at several target-to-masker ratios. For the normal-hearing listeners, target-masker combinations were processed through a noise-excited vocoder designed to simulate a cochlear implant. With unprocessed stimuli, a normal-hearing control group maintained high levels of intelligibility down to target-to-masker ratios as low as 0 dB and showed a release from masking, producing better performance with single-talker maskers than with steady-state noise. In contrast, no masking release was observed in either implant or normal-hearing subjects listening through an implant simulation. The performance of the simulation and implant groups did not improve when the single-talker masker was a different talker compared to the same talker as the target speech, as was found in the normal-hearing control. These results are interpreted as evidence for a significant role of informational masking and modulation interference in cochlear implant speech recognition with fluctuating maskers. This informational masking may originate from increased target-masker similarity when spectral resolution is reduced.     

Speech intelligibility and localization in a multi-source environment.

Natural environments typically contain sound sources other than the source of interest that may interfere with the ability of listeners to extract information about the primary source. Studies of speech intelligibility and localization by normal-hearing listeners in the presence of competing speech are reported on in this work. One, two or three competing sentences [IEEE Trans. Audio Electroacoust. 17(3), 225-246 (1969)] were presented from various locations in the horizontal plane in several spatial configurations relative to a target sentence. Target and competing sentences were spoken by the same male talker and at the same level. All experiments were conducted both in an actual sound field and in a virtual sound field. In the virtual sound field, both binaural and monaural conditions were tested. In the speech intelligibility experiment, there were significant improvements in performance when the target and competing sentences were spatially separated. Performance was similar in the actual sound-field and virtual sound-field binaural listening conditions for speech intelligibility. Although most of these improvements are evident monaurally when using the better ear, binaural listening was necessary for large improvements in some situations. In the localization experiment, target source identification was measured in a seven-alternative absolute identification paradigm with the same competing sentence configurations as for the speech study. Performance in the localization experiment was significantly better in the actual sound-field than in the virtual sound-field binaural listening conditions. Under binaural conditions, localization performance was very good, even in the presence of three competing sentences. Under monaural conditions, performance was much worse. For the localization experiment, there was no significant effect of the number or configuration of the competing sentences tested. For these experiments, the performance in the speech intelligibility experiment was not limited by localization ability.     

Multichannel speech intelligibility and talker recognition using monaural, binaural, and three-dimensional auditory presentation

In a 3D auditory display, sounds are presented over headphones in a way that they seem to originate from virtual sources in a space around the listener. This paper describes a study on the possible merits of such a display for bandlimited speech with respect to intelligibility and talker recognition against a background of competing voices. Different conditions were investigated: speech material (words/sentences), presentation mode (monaural/binaural/3D), number of competing talkers (1-4), and virtual position of the talkers (in 45 degrees-steps around the front horizontal plane). Average results for 12 listeners show an increase of speech intelligibility for 3D presentation for two or more competing talkers compared to conventional binaural presentation. The ability to recognize a talker is slightly better and the time required for recognition is significantly shorter for 3D presentation in the presence of two or three competing talkers. Although absolute localization of a talker is rather poor, spatial separation appears to have a significant effect on communication. For either speech intelligibility, talker recognition, or localization, no difference is found between the use of an individualized 3D auditory display and a general display.     

Improved speech recognition in noise in simulated binaurally combined acoustic and electric stimulation

Speech recognition in noise improves with combined acoustic and electric stimulation compared to electric stimulation alone [Kong et al., J. Acoust. Soc. Am. 117, 1351-1361 (2005)]. Here the contribution of fundamental frequency (F0) and low-frequency phonetic cues to speech recognition in combined hearing was investigated. Normal-hearing listeners heard vocoded speech in one ear and low-pass (LP) filtered speech in the other. Three listening conditions (vocode-alone, LP-alone, combined) were investigated. Target speech (average F0=120 Hz) was mixed with a time-reversed masker (average F0=172 Hz) at three signal-to-noise ratios (SNRs). LP speech aided performance at all SNRs. Low-frequency phonetic cues were then removed by replacing the LP speech with a LP equal-amplitude harmonic complex, frequency and amplitude modulated by the F0 and temporal envelope of voiced segments of the target. The combined hearing advantage disappeared at 10 and 15 dB SNR, but persisted at 5 dB SNR. A similar finding occurred when, additionally, F0 contour cues were removed. These results are consistent with a role for low-frequency phonetic cues, but not with a combination of F0 information between the two ears. The enhanced performance at 5 dB SNR with F0 contour cues absent suggests that voicing or glimpsing cues may be responsible for the combined hearing benefit.     

Effects of cross-language voice training on speech perception: whose familiar voices are more intelligible?

Previous research has shown that familiarity with a talker's voice can improve linguistic processing (herein, "Familiar Talker Advantage"), but this benefit is constrained by the context in which the talker's voice is familiar. The current study examined how familiarity affects intelligibility by manipulating the type of talker information available to listeners. One group of listeners learned to identify bilingual talkers' voices from English words, where they learned language-specific talker information. A second group of listeners learned the same talkers from German words, and thus only learned language-independent talker information. After voice training, both groups of listeners completed a word recognition task with English words produced by both familiar and unfamiliar talkers. Results revealed that English-trained listeners perceived more phonemes correct for familiar than unfamiliar talkers, while German-trained listeners did not show improved intelligibility for familiar talkers. The absence of a processing advantage in speech intelligibility for the German-trained listeners demonstrates limitations on the Familiar Talker Advantage, which crucially depends on the language context in which the talkers' voices were learned; knowledge of how a talker produces linguistically relevant contrasts in a particular language is necessary to increase speech intelligibility for words produced by familiar talkers.     

Speech perception and talker segregation: effects of level, pitch, and tactile support with multiple simultaneous talkers

Speech intelligibility was investigated by varying the number of interfering talkers, level, and mean pitch differences between target and interfering speech, and the presence of tactile support. In a first experiment the speech-reception threshold (SRT) for sentences was measured for a male talker against a background of one to eight interfering male talkers or speech noise. Speech was presented diotically and vibro-tactile support was given by presenting the low-pass-filtered signal (0-200 Hz) to the index finger. The benefit in the SRT resulting from tactile support ranged from 0 to 2.4 dB and was largest for one or two interfering talkers. A second experiment focused on masking effects of one interfering talker. The interference was the target talker's own voice with an increased mean pitch by 2, 4, 8, or 12 semitones. Level differences between target and interfering speech ranged from -16 to +4 dB. Results from measurements of correctly perceived words in sentences show an intelligibility increase of up to 27% due to tactile support. Performance gradually improves with increasing pitch difference. Louder target speech generally helps perception, but results for level differences are considerably dependent on pitch differences. Differences in performance between noise and speech maskers and between speech maskers with various mean pitches are explained by the effect of informational masking.     

Spatial release from masking in normally hearing and hearing-impaired listeners as a function of the temporal overlap of competing talkers

Listeners with sensorineural hearing loss are poorer than listeners with normal hearing at understanding one talker in the presence of another. This deficit is more pronounced when competing talkers are spatially separated, implying a reduced "spatial benefit" in hearing-impaired listeners. This study tested the hypothesis that this deficit is due to increased masking specifically during the simultaneous portions of competing speech signals. Monosyllabic words were compressed to a uniform duration and concatenated to create target and masker sentences with three levels of temporal overlap: 0% (non-overlapping in time), 50% (partially overlapping), or 100% (completely overlapping). Listeners with hearing loss performed particularly poorly in the 100% overlap condition, consistent with the idea that simultaneous speech sounds are most problematic for these listeners. However, spatial release from masking was reduced in all overlap conditions, suggesting that increased masking during periods of temporal overlap is only one factor limiting spatial unmasking in hearing-impaired listeners.     

Acoustic and linguistic factors in the perception of bandpass-filtered speech

Speech can remain intelligible for listeners with normal hearing when processed by narrow bandpass filters that transmit only a small fraction of the audible spectrum. Two experiments investigated the basis for the high intelligibility of narrowband speech. Experiment 1 confirmed reports that everyday English sentences can be recognized accurately (82%-98% words correct) when filtered at center frequencies of 1500, 2100, and 3000 Hz. However, narrowband low predictability (LP) sentences were less accurately recognized than high predictability (HP) sentences (20% lower scores), and excised narrowband words were even less intelligible than LP sentences (a further 23% drop). While experiment 1 revealed similar levels of performance for narrowband and broadband sentences at conversational speech levels, experiment 2 showed that speech reception thresholds were substantially (>30 dB) poorer for narrowband sentences. One explanation for this increased disparity between narrowband and broadband speech at threshold (compared to conversational speech levels) is that spectral components in the sloping transition bands of the filters provide important cues for the recognition of narrowband speech, but these components become inaudible as the signal level is reduced. Experiment 2 also showed that performance was degraded by the introduction of a speech masker (a single competing talker). The elevation in threshold was similar for narrowband and broadband speech (11 dB, on average), but because the narrowband sentences required considerably higher sound levels to reach their thresholds in quiet compared to broadband sentences, their target-to-masker ratios were very different (+23 dB for narrowband sentences and -12 dB for broadband sentences). As in experiment 1, performance was better for HP than LP sentences. The LP-HP difference was larger for narrowband than broadband sentences, suggesting that context provides greater benefits when speech is distorted by narrow bandpass filtering.     

The interlanguage speech intelligibility benefit

This study investigated how native language background influences the intelligibility of speech by non-native talkers for non-native listeners from either the same or a different native language background as the talker. Native talkers of Chinese (n = 2), Korean (n = 2), and English (n = 1) were recorded reading simple English sentences. Native listeners of English (n = 21), Chinese (n = 21), Korean (n = 10), and a mixed group from various native language backgrounds (n = 12) then performed a sentence recognition task with the recordings from the five talkers. Results showed that for native English listeners, the native English talker was most intelligible. However, for non-native listeners, speech from a relatively high proficiency non-native talker from the same native language background was as intelligible as speech from a native talker, giving rise to the "matched interlanguage speech intelligibility benefit." Furthermore, this interlanguage intelligibility benefit extended to the situation where the non-native talker and listeners came from different language backgrounds, giving rise to the "mismatched interlanguage speech intelligibility benefit." These findings shed light on the nature of the talker-listener interaction during speech communication.     

The effect of segmental and suprasegmental corrections on the intelligibility of deaf speech

Three experiments were conducted to study the effect of segmental and suprasegmental corrections on the intelligibility and judged quality of deaf speech. By means of digital signal processing techniques, including LPC analysis, transformations of separate speech sounds, temporal structure, and intonation were carried out on 30 Dutch sentences spoken by ten deaf children. The transformed sentences were tested for intelligibility and acceptability by presenting them to inexperienced listeners. In experiment 1, LPC based reflection coefficients describing segmental characteristics of deaf speakers were replaced by those of hearing speakers. A complete segmental correction caused a dramatic increase in intelligibility from 24% to 72%, which, for a major part, was due to correction of vowels. Experiment 2 revealed that correction of temporal structure and intonation caused only a small improvement from 24% to about 34%. Combination of segmental and suprasegmental corrections yielded almost perfectly understandable sentences, due to a more than additive effect of the two corrections. Quality judgments, collected in experiment 3, were in close agreement with the intelligibility measures. The results show that, in order for these speakers to become more intelligible, improving their articulation is more important than improving their production of temporal structure and intonation.     

Contributions of talker characteristics and spatial location to auditory streaming

To examine whether auditory streaming contributes to unmasking, intelligibility of target sentences against two competing talkers was measured using the coordinate response measure (CRM) [Bolia et al., J. Acoust. Soc. Am. 107, 1065-1066 (2007)] corpus. In the control condition, the speech reception threshold (50% correct) was measured when the target and two maskers were collocated straight ahead. Separating maskers from the target by +/-30 degrees resulted in spatial release from masking of 12 dB. CRM sentences involve an identifier in the first part and two target words in the second part. In experimental conditions, masking talkers started spatially separated at +/-30 degrees but became collocated with the target before the scoring words. In one experiment, one target and two different maskers were randomly selected from a mixed-sex corpus. Significant unmasking of 4 dB remained despite the absence of persistent location cues. When same-sex talkers were used as maskers and target, unmasking was reduced. These data suggest that initial separation may permit confident identification and streaming of the target and masker speech where significant differences between target and masker voice characteristics exist, but where target and masker characteristics are similar, listeners must rely more heavily on continuing spatial cues.     

The clear speech effect for non-native listeners

Previous work has established that naturally produced clear speech is more intelligible than conversational speech for adult hearing-impaired listeners and normal-hearing listeners under degraded listening conditions. The major goal of the present study was to investigate the extent to which naturally produced clear speech is an effective intelligibility enhancement strategy for non-native listeners. Thirty-two non-native and 32 native listeners were presented with naturally produced English sentences. Factors that varied were speaking style (conversational versus clear), signal-to-noise ratio (-4 versus -8 dB) and talker (one male versus one female). Results showed that while native listeners derived a substantial benefit from naturally produced clear speech (an improvement of about 16 rau units on a keyword-correct count), non-native listeners exhibited only a small clear speech effect (an improvement of only 5 rau units). This relatively small clear speech effect for non-native listeners is interpreted as a consequence of the fact that clear speech is essentially native-listener oriented, and therefore is only beneficial to listeners with extensive experience with the sound structure of the target language.     

Intelligibility ratings of continuous discourse: application to hearing aid selection

Twelve normal-hearing subjects rated the intelligibility of 35-s, hearing-aid-processed continuous discourse (CD) passages. Three talkers (two male, one female), four hearing aids, and two signal-to-babble (S/B) ratios were used in a completely crossed design. Research questions concerned: (1) ability of listeners to rate intelligibility, (2) sensitivity of hearing aid rankings were based on intelligibility ratings for three CD passages per instrument, and (3) dependence of hearing aid rankings on (a) S/B ratio, and (b) talker characteristics. Results were: (1) listeners were able to rate intelligibility, (2) rankings based on intelligibility ratings of three CD passages per hearing aid were capable of identifying two superior instruments within a group of four hearing aids that were similar in frequency/gain function, (3) listening in a more difficult S/B ratio substantially decreased the sensitivity of the hearing aid rankings for the female talker but had only minor effects on the rankings for the male talkers, and (4) hearing aid intelligibility rankings were found to be different for different talkers. Applications to hearing aid selection are discussed.     

Age-related changes in talker recognition with reduced spectral cues

Temporal information provided by cochlear implants enables successful speech perception in quiet, but limited spectral information precludes comparable success in voice perception. Talker identification and speech decoding by young hearing children (5-7 yr), older hearing children (10-12 yr), and hearing adults were examined by means of vocoder simulations of cochlear implant processing. In Experiment 1, listeners heard vocoder simulations of sentences from a man, woman, and girl and were required to identify the talker from a closed set. Younger children identified talkers more poorly than older listeners, but all age groups showed similar benefit from increased spectral information. In Experiment 2, children and adults provided verbatim repetition of vocoded sentences from the same talkers. The youngest children had more difficulty than older listeners, but all age groups showed comparable benefit from increasing spectral resolution. At comparable levels of spectral degradation, performance on the open-set task of speech decoding was considerably more accurate than on the closed-set task of talker identification. Hearing children's ability to identify talkers and decode speech from spectrally degraded material sheds light on the difficulty of these domains for child implant users.