Target Voice Research Articles

Segment imitation in unfamiliar language varieties

Phonetic imitation has been explored in various research: in studies on impersonation [Zetterholm (2001); spontaneous phonetic imitation [Babel (2012)]; and as part of studies investigating overall imitation [Papcun (1988), Babel (2011)]. There are fewer studies focusing on specific phonemes, and the context required for successful imitation. Babel [186 (2012)] (Am. English) and Zetterholm [281 (1997)] (Swedish) found spontaneous imitation was more common in the open vowels than in close. A case study by Zetterholm [275 (1997)] of one impersonator shows formant measurements as closer to the target voice than that of the impersonator. This paper presents a comparative acoustic analysis of imitation of unfamiliar phonemes by untrained imitators, based on a small study containing recordings from SWEDIA and SUF. Vowel tokens are edited to three lengths with varying degrees of context. Listeners are asked to imitate the sound, and attempt to use it in a word. Acoustic analysis of F0, 1, 2, and 3 will be compared to the original recordings as well as the speakers’ own speech. The conclusion and discussion includes indications of how easy it is to produce phonemes when they are not native to a speaker’s own variant, as well as its relation to the forensic context.

Multi-scale Speaker Transformation Using Radial Basis Function

In speaker transformation, the speaker dependent spectral parameters are generally characterized by single scale features. These features approximate the vocal tract, but produce artifacts during speech signal reconstruction. In this paper, multi-resolution wavelet based feature set is proposed, which finely tunes the speaker specific characteristics of the speech signal. The Radial Basis Function is used to propose the mapping function for modifying these characteristics. The performance of the proposed system is evaluated using different objective and subjective measures. Evaluation results illustrate that the proposed algorithm maintains target voice individuality while maintaining the quality and naturalness of the speech signal.

Selective Speech Feature Extraction using Channel Similarity in CHMM Vocabulary Recognition

HMM Speech recognition systems have a few weaknesses, including failure to recognize speech due to the mixing of environment noise other voices. In this paper, we propose a speech feature extraction methode using CHMM for extracting selected target voice from mixture of voices and noises. we make use of channel similarity and correlate relation for the selective speech extraction composes. This proposed method was validated by showing that the average distortion of separation of the technique decreased by 0.430 dB. It was shown that the performance of the selective feature extraction is better than another system.

Measurement of Acceptable Noise Level and Self-Reported Subjective Disability in Adult Cochlear Implant Users

The present study aimed to compare the acceptable noise levels (ANLs) of normal-hearing listeners (N = 10) and cochlear implant (CI, N = 10) users, depending on the target-talker gender and the number of background competing talkers. The profile of hearing disability of CI users was also determined by the Korean-translated speech, spatial, and qualities of hearing scale (K-SSQ) for CI users. Results showed that the ANLs did not significantly differ between CI users and normal-hearing listeners. For both groups, the effect of target-talker gender (male vs. female) was not significant, whereas the number of background competing talkers (1 to 8 talkers) significantly influenced ANL results. The ANLs of both groups significantly increased as the number of competing talkers was reduced. The greatest distraction from noise was observed when the gender of target and background talkers was matched. This suggests that, regardless of listener groups, all the listeners are more annoyed by background noise when the background speech becomes more meaningful or when the voice of target and competing talkers gets similar. Correlational analyses revealed that the ANLs of CI users were significantly related to the K-SSQ responses on the quality of hearing domain, indicating that CI users who accepted more amount of noise actually reported more improvement on sound quality with CI use.

Voice line-ups: Speakers’ F0 values influence the reliability of voice recognitions

This paper reports a voice line-up experiment. It was designed to explore whether some voices are easier to remember and recognise than others, and whether some listeners are better at recognising voices than others. The results suggest that 'common' voices, based on mean fundamental frequency values, are harder to remember and recognise than 'less common' voices. The results also suggest that some people are better at remembering and recognising voices than others. The results show that after one week of retention the mean percentage of correct identifications is 65%. However, the 'less common' voice is correctly recognised at 74%, whereas the 'common' voice is only correctly identified in 56% of the cases. The results not only suggest that some voices are easier to remember and recognise than others, but they also imply individual differences in listeners' abilities to recognise voices. It turns out that it was the same listeners who had difficulties making the correct identifications in the different line-ups. An implicational hierarchy may be proposed, in which the listeners who were not able to correctly identify the 'less common' voice also had problems with identifying the 'common' voice. The results also show that the listeners' confidence in their own ability to recognise the target voices is not a reliable measure of voice recognition performance.

Earwitnesses: the effect of type of vocal differences on correct identification and confidence accuracy

Individual voice and speech characteristics are important for earwitness identification. A target-absent line-up with six foils was used to analyse the influence of voice and speech features on recognition. To create ecologically valid conditions for voice recognition, the present study used voices in conversation in the target event and for the foils. The participants were particularly successful in rejecting the two voice foils that were most dissimilar to the target voice in articulation rate and pitch. These two foils were also given somewhat higher confidence judgments than the other foils, although the level of that confidence was still somewhat low. In fact, participants as a collective were underconfident for these two foils but were overconfident for the other foils. In addition, they showed somewhat better ability in their confidence judgments to distinguish correct and incorrect responses for the two most dissimilar foils than for the four foils for which their identification responses were less successful. For the other four foils, which were more similar to the target voice, the participants showed in their confidence judgments very poor ability to distinguish between correct and incorrect identifications. Later positions in the parade were more often (erroneously) identified as the culprit, regardless of the specific foils used in those positions.

The Benefit of Bilateral Versus Unilateral Cochlear Implantation to Speech Intelligibility in Noise

To develop a predictive model of spatial release from masking (SRM) for cochlear implantees, and validate this model against data from the literature. To establish the spatial configurations for which the model predicts a large advantage of bilateral over unilateral implantation. To collect data to support these predictions and generate predictions of more typical advantages of bilateral implantation. The model initially assumed that bilateral cochlear implantees had equally effective implants on each side, with which they could perform optimal better-ear listening. Predictions were compared with measurements of SRM, using one and two implants with up to three interfering noises. The effect of relaxing the assumption of equally effective implants was explored. Novel measurements of SRM for eight unilateral implantees were collected, including measurements using speech and noise at azimuths of ± 60 degrees, and compared with prediction. A spatial map of bilateral implant benefit was generated for a situation with one interfering noise in anechoic conditions, and predictions of benefit were generated from binaural room impulse responses in a variety of real rooms. The model accurately predicted data from a previous study for multiple interfering noises in a variety of spatial configurations, even when implants were assumed to be equally effective (r = 0.97). It predicted that the maximum benefit of bilateral implantation was 18 dB. Predictions were little affected if the implants were not assumed to be equally effective. The new measurements supported the 18 dB advantage prediction. The spatial map of predicted benefit showed that, for a listener facing the target voice, bilateral implantees could enjoy an advantage of about 10 dB over unilateral implantees in a wide range of situations. Predictions based on real-room measurements with speech and noise at 1 m showed that large benefits can occur even in reverberant spaces. In optimal conditions, the benefit of bilateral implantation to speech intelligibility in noise can be much larger than has previously been reported. This benefit is thus considerably larger than reported benefits of summation or squelch and is robust in reverberation when the interfering source is close.

Recognition and description of singing voices: The impact of verbal overshadowing

Verbalizing sound quality of individual music performers presents a challenge to musicians and pedagogues in describing a complex aural phenomenon. Verbal overshadowing (VO) is well documented in psychological literature and can occur when we use words to describe sensory experiences (such as seeing, tasting or hearing). The verbal description impairs later recall of the sensory experience and individuals are less able to identify the original from a line-up of similar stimuli. This study investigated the impact of verbal description on listeners’ recognition accuracy and confidence of classical singing voices. It also assessed the scope and limitations of the terms used to describe singers’ sound. Five soprano singers performed an excerpt of Caro mio ben twice for a recording. Listeners (n = 50) heard a single singer performing the first take of Caro mio ben, which they either described in words, or just remembered. Listeners were later asked to identify the voice from a homogenous line-up of the five singers. Providing a verbal description reduced listeners’ identification accuracy of target voice, but did not reduce listeners’ confidence in their decisions. There was a reliable verbal overshadowing effect and results suggest that verbalization of singing voices is problematic and limits listeners’ ability to remember a singer’s unique sound quality.

무음 압축을 이용하는 음성 통신 시스템을 위한 동적 버퍼 관리 알고리즘

본 논문은 무음 압축(silence suppression)을 이용하는 음성 코덱 이용 시 음성 통신 용량(voice capacity)을 늘리기 위한 동적 버퍼 관리 알고리즘인 silence drop first(SDF) 알고리즘을 제안한다. SDF 알고리즘은 음성 통신을 위한 버퍼에서 오버플로우가 발생할 경우 무음 패킷을 음성 패킷보다 먼저 폐기하는 기법을 이용하여 상대적으로 더 많은 음성 패킷을 종단간 전송한다. 알고리즘의 성능 평가를 위하여 G.729A 및 G.711 음성 코덱을 사용하였다. 성능 분석 결과 제안하는 기법을 이용할 경우 MOS가 3.0인 수준의 품질을 유지하면서 G.729A 코덱에 대해 약 84.21%, G.711 코덱에 대해 약 38.46%의 사용자가 추가로 네트워크 이용이 가능함을 확인할 수 있었다. 또한 제안하는 기법 이용 시 동일 사용자 수에 대해 음성 통화 품질을 보장하기 위해 필요한 무음 패킷생성 주기 및 링크 용량의 제한을 완화할 수 있음을 확인하였다. This paper proposes silence drop first(SDF) active buffer management algorithm to increase the voice capacity when silence suppression is used. This algorithm finds and drops silence packet rather than voice packet in the queue for resolving buffer overflow of queue. Simulations with voice codec of G.729A and G.711 are performed. By using proposed SDF algorithm, the voice capacity is increased by 84.21% with G.729A and 38.46% with G.711. Further more, SDF algorithm reduces the required link capacity and loosens the silence packet inter-arrival time limit to provide target voice quality compared with that of conventional algorithms.

Employing an active mental task to enhance the performance of auditory attention-based brain–computer interfaces

ObjectiveA majority of auditory brain–computer interfaces (BCIs) use the attentional modulation of auditory event-related potentials (ERPs) for communication and control. This study investigated whether the performance of an ERP-based auditory BCI can be further improved by increasing the mental efforts associated with the execution of the attention-related task. MethodsSubjects mentally selected a target among a random sequence of spoken digits. Upon the detection of the target digit, the subjects were required to perform an active mental task (AMT) – mentally discriminating the gender property of the target voice. The total number of presented digits was manipulated to investigate possible influences of the number of choices. The subjects also participated in two control experiments, in which they were asked to (1) press a button to report their discrimination results or (2) simply count the appearance of the target digit without performing the AMT. ResultsTwo ERP components, that is, a negative shift around 200ms (Nd) over the fronto-central area and a positive deflection during 500–600ms (late positive component, LPC) over the central–parietal area, were modulated by execution of the AMT. Compared to a counting task, the AMT resulted in paradigm-specific enhanced LPC responses. The latency of the LPC was significantly correlated with the behavioural reaction time, indicating that the LPC could originate from a response-related brain network similar to P3b. The AMT paradigm resulted in an increase of 4–6% in BCI classification accuracies, compared to a counting paradigm that was considered to represent the traditional auditory attention BCI paradigms (p<0.05). In addition, the BCI classification accuracies were not significantly affected by the number of BCI choices in the AMT paradigm. Conclusions(1) LPC was identified as the AMT-specific ERP component and (2) the performance of auditory BCIs can be improved from the human response side by introducing additional mental efforts when executing attention-related tasks. SignificanceThe neurophysiological characteristics of the recently proposed auditory BCI paradigm using an AMT were explored. The results suggest the proposed paradigm as a candidate for improving the performance of auditory BCIs.

Voice gender differences and separation of simultaneous talkers in cochlear implant users with residual hearing

Perception of a target voice in the presence of a competing talker, of same or different gender as the target, was investigated in cochlear implant users, in implant-alone and bimodal (acoustic hearing in the non-implanted ear) conditions. Recordings of two male and two female talkers acted as targets and maskers, to investigate whether bimodal benefit increased for different compared to same gender target/maskers due to increased ability to perceive and utilize fundamental frequency and spectral-shape differences. In both listening conditions participants showed benefit of target/masker gender difference. There was an overall bimodal benefit, which was independent of target/masker gender difference.

Predictions of spatial release from masking from architectural plans

Spatial separation of target speech and interfering noise produces spatial release from masking (SRM). A model of SRM for speech in noise gives accurate predictions for multiple noise sources and in reverberation. The model is based on the frequency-weighted combination of better ear listening and binaural unmasking. Here we use this model to explore the effects of room acoustics, seating choice, listener head orientation and table layout in a virtual restaurant simulation. The modeled restaurant contains nine tables for two in three rows, and each diner listens to their partner across the table. Substantial SRM was predicted in all seating locations, but was lowest at the centre table. Orienting the head away from the target voice by 20-30deg improved SRM for most seating positions, but this benefit was small for those seated at the edges of the restaurant and facing in. Reorienting these tables so that they have a wall at one side improved average SRM. Acoustic treatments applied to the walls produced larger benefits in SRM than treatment of the ceiling that achieved equivalent reverberation time. Reverberation reduced both SRM and its variability across seating locations and head orientations.

Pitch strength of normal and dysphonic voices

Two sounds with the same pitch may vary from each other based on saliency of their pitch sensation. This perceptual attribute is called "pitch strength." The study of voice pitch strength may be important in quantifying of normal and pathological qualities. The present study investigated how pitch strength varies across normal and dysphonic voices. A set of voices (vowel /a/) selected from the Kay Elemetrics Disordered Voice Database served as the stimuli. These stimuli demonstrated a wide range of voice quality. Ten listeners judged the pitch strength of these stimuli in an anchored magnitude estimation task. On a given trial, listeners heard three different stimuli. The first stimulus represented very low pitch strength (wide-band noise), the second stimulus consisted of the target voice and the third stimulus represented very high pitch strength (pure tone). Listeners estimated pitch strength of the target voice by positioning a continuous slider labeled with values between 0 and 1, reflecting the two anchor stimuli. Results revealed that listeners can judge pitch strength reliably in dysphonic voices. Moderate to high correlations with perceptual judgments of voice quality suggest that pitch strength may contribute to voice quality judgments.

Voice Conversion Using Dynamic Kernel Partial Least Squares Regression

A drawback of many voice conversion algorithms is that they rely on linear models and/or require a lot of tuning. In addition, many of them ignore the inherent time-dependency between speech features. To address these issues, we propose to use dynamic kernel partial least squares (DKPLS) technique to model nonlinearities as well as to capture the dynamics in the data. The method is based on a kernel transformation of the source features to allow non-linear modeling and concatenation of previous and next frames to model the dynamics. Partial least squares regression is used to find a conversion function that does not overfit to the data. The resulting DKPLS algorithm is a simple and efficient algorithm and does not require massive tuning. Existing statistical methods proposed for voice conversion are able to produce good similarity between the original and the converted target voices but the quality is usually degraded. The experiments conducted on a variety of conversion pairs show that DKPLS, being a statistical method, enables successful identity conversion while achieving a major improvement in the quality scores compared to the state-of-the-art Gaussian mixture-based model. In addition to enabling better spectral feature transformation, quality is further improved when aperiodicity and binary voicing values are converted using DKPLS with auxiliary information from spectral features.

Voice segregation by difference in fundamental frequency: Evidence for harmonic cancellation

Two experiments investigated listeners' ability to use a difference of two semitones in fundamental frequency (F0) to segregate a target voice from harmonic complex tones, with speech-like spectral profiles. Masker partials were in random phase (experiment 1) or in sine phase (experiment 2) and stimuli were presented over headphones. Target's and masker's harmonicity were each distorted by F0 modulation and reverberation. The F0 of each source was manipulated (monotonized or modulated by 2 semitones at 5 Hz) factorially. In addition, all sources were presented from the same location in a virtual room with controlled reverberation, assigned factorially to each source. In both experiments, speech reception thresholds increased by about 2 dB when the F0 of the masker was modulated and increased by about 6 dB when, in addition to F0 modulation, the masker was reverberant. Masker partial phases did not influence the results. The results suggest that F0-segregation relies upon the masker's harmonicity, which is disrupted by rapid modulation. This effect is compounded by reverberation. In addition, F0-segregation was found to be independent of the depth of masker envelope modulations.

Executive abilities for spoken-word commands: Inhibiting conflicting responses in voice-tone classification by adolescents and adults

The inhibition of responses prompted by conflicting laterality cues and response-mapping rules was studied while participants classified the voice-tone of word commands. Announced by the cue words /left/ or /right/, target word commands for impeding or instigating actions (e.g., /quit/ or /go/), spoken in stern or lenient tone, stimulated the left or right ear. Each trial presented a cue followed by a target (e.g., left-quit), and participants classified the target voice tone as lenient or stern with a left or right (or the reverse) response for impeding (or instigating) commands. Within a trial block, laterality-conflict conditions presented impeding or instigating words only and, on each trial, the cue and ear side were congruent or in conflict with the correct response side. Response-mapping conflict conditions presented impeding and instigating words within each trial block, and correct classification required adhering to the respective mapping. Without conflict, error percent was larger with instigating than impeding commands, and in adolescents than adults. With instigating commands, laterality conflict increased the errors in both groups but only in adolescents with impeding commands. Errors increased further with laterality and mapping conflict, especially in adolescents. The adolescent response inhibition is inferior to the adult. [Funded by ABMRF.]

Remembering, Recognizing and Describing Singers' Sound Identities

Perceiving music performers through sound quality is important for musicians, teachers and every music listener. A performer's musical identity is influenced by the unique sound they create, yet little is known about listeners' mental representation of an individual's sound quality. A verbal overshadowing (VO) effect has been extensively reported in visual recognition, where the act of verbal description distorts subsequent recall, and the same may apply in the auditory domain and specifically, in music performance reception. In this study we explore if musicians also exhibit a VO effect, by investigating how listeners form auditory impressions of a singer and testing the impact of verbal description on listeners' recognition accuracy and confidence. Listeners (n = 50) heard a single target voice and were assigned to either a verbal description group or a control group. The verbal group wrote a detailed description of the target voice and the control group did not. All listeners heard a voice line-up (n = 6) and were asked to identify the original target voice and rate their confidence in the decision. There was a reliable VO effect where providing a verbal description reduced listeners' identification accuracy of the target voice, but did not reduce listeners' confidence rating of their decision. Verbalization of singing voices was an inadequate representation of listeners' auditory experience and limited listeners' ability to recall and identify singers' sound quality. These results have profound implications for the way in which we consider musical performance and how music listeners recognize, process and describe a performer's sound.

Typicality effects on memory for voice: Implications for earwitness testimony

AbstractThe effects of voice typicality on voice recognition were examined. In an incidental task, participants were presented with consonant‐vowel‐consonant stimuli (with varying vowels) spoken by voices rated as high‐typical or low‐typical. One week later, they were brought back for a surprise voice recognition test where the target voice was presented along with low‐typical and high‐typical voice distracters. The results indicated that voice misidentifications occur, with many errors due to confusions of high‐typical voice targets with high‐typical target voices. The results show that voice typicality affects voice recognition accuracy and is an important factor to be considered in real‐life earwitness testimony situations. Copyright © 2009 John Wiley &amp; Sons, Ltd.

Method And System For The Quick Conversion Of A Voice Signal

A method for converting a voice signal from a source speaker into a converted voice signal with acoustic characteristics similar to those of a target speaker includes the steps of determining (1) at least one function for transforming source speaker acoustic characteristics into acoustic characteristics similar to those of the target speaker using target and source speaker voice samples; and transforming acoustic characteristics of the source speaker voice signal to be converted by applying said transformation function(s). The method is characterised in that said transformation (2) includes the step (44) of applying only a predetermined portion of at least one transformation function to said signal to be converted.

2-Year-Olds' Speech Understanding in Multitalker Environments.

Infants and toddlers are often spoken to in the presence of background sounds, including speech from other talkers. Prior work has suggested that infants 1 year of age and younger can only recognize speech when it is louder than any distracters in the environment. The present study tests 24-month-olds' ability to understand speech in a multitalker environment. Children were presented with a preferential-looking task in which a target voice told them to find one of two objects. At the same time, multitalker babble was presented as a distracter, at one of four signal-to-noise ratios. Children showed some ability to understand speech and look at the appropriate referent at signal-to-noise ratios as low as -5 dB. These findings suggest that 24-month-olds are better able to selectively attend to an interesting voice in the context of competing distracter voices than are younger infants. There were significant correlations between individual children's performance and their vocabulary size, but only at one of the four noise levels; thus, it does not appear that vocabulary size is the driving factor in children's listening improvement, although it may be a contributing factor to performance in noisy environments.