Typical Speaker Research Articles

Version control of speaker recognition systems

This paper discusses one of the most challenging practical engineering problems in speaker recognition systems — the version control of models and user profiles. A typical speaker recognition system consists of two stages: the enrollment stage, where a profile is generated from user-provided enrollment audio; and the runtime stage, where the voice identity of the runtime audio is compared against the stored profiles. As technology advances, the speaker recognition system needs to be updated for better performance. However, if the stored user profiles are not updated accordingly, version mismatch will result in meaningless recognition results. In this paper, we describe different version control strategies for speaker recognition systems that had been carefully studied at Google from years of engineering practice. These strategies are categorized into three groups according to how they are deployed in the production environment: device-side deployment, server-side deployment, and hybrid deployment. To compare different strategies with quantitative metrics under various network configurations, we present SpeakerVerSim, an easily-extensible Python-based simulation framework for different server-side deployment strategies of speaker recognition systems.

When Speaker Recognition Meets Noisy Labels: Optimizations for Front-Ends and Back-Ends

A typical speaker recognition system often involves two modules: a feature extractor front-end and a speaker identification back-end. Despite the superior performance that deep neural networks have achieved for the front-end, their success benefits from the availability of large-scale, correctly labeled datasets. While label noise is unavoidable in speaker recognition datasets, both the front-end and back-end are affected by label noise, which degrades speaker recognition performance. In this paper, we first conduct comprehensive experiments to help improve our understanding of the effects of label noise on both the front-end and back-end. Then, we propose a simple yet effective training paradigm and loss correction method to handle label noise in the front-end. We combine our proposed method with the recently proposed Bayesian estimation of PLDA for noisy labels, and the whole system shows strong robustness to label noise. Furthermore, we show two practical applications of the improved system: one application corrects noisy labels based on an utterance’s chunk-level predictions, and the other algorithmically filters out high-confidence noisy samples within a dataset. By applying the second application to the NIST SRE04–10 dataset and verifying filtered utterances by human validation, we identify that approximately 1% of the NIST SRE04–10 dataset is made up of label errors.

Behavior Assessment Battery: Normative and Psychometric Investigation Among Polish Adults Who Do and Do Not Stutter.

The aim of this study was to establish normative data for Polish persons who stutter (PWS) and persons who do not stutter (PWNS) with the adapted version of the Behavior Assessment Battery (BAB; Vanryckeghem & Brutten, 2018) for adults and to investigate whether or not gender and age influenced the results. In addition, the BAB subtests' reliability and validity were examined. A Polish version of the BAB test procedures, including the Speech Situation Checklist-Emotional Reaction, Speech Situation Checklist-Speech Disruption, Behavior Checklist, and Communication Attitude Test for Adults Who Stutter (Vanryckeghem & Brutten, 2018), were individually administered to 123 adults who stutter and 151 adults who do not stutter between the ages of 17 and 70. PWS scored statistically significantly higher on all BAB subtests compared to PWNS. No significant between-groups gender or age difference was detected. The test data indicate high internal validity. The test items correlate well with the total score and significantly differentiate the two groups. The test battery validly differentiates PWS from PWNS with high accuracy. The Polish version of the BAB for adults confirms data collected in previous cross-cultural investigations. The results once again differentiate PWS from PWNS in a powerful way. Compared to PWNS, PWS report significantly more anxiety and speech breakdown when speaking in particular speech situations. They also report their attitude toward speech to be significantly more negative compared to the typical speaker and indicate using a significant number of coping behaviors to avoid or escape difficult speaking conditions. The BAB test battery, in its totality, assists in multidimensional assessment, pre- and posttreatment comparison, and identification of treatment targets based on the answers of PWS to each test item.

What does it take to stress a word? Digital manipulation of stress markers in ataxic dysarthria.

Stress production is important for effective communication, but this skill is frequently impaired in people with motor speech disorders. The literature reports successful treatment of these deficits in this population, thus highlighting the therapeutic potential of this area. However, no specific guidance is currently available to clinicians about whether any of the stress markers are more effective than others, to what degree they have to be manipulated, and whether strategies need to differ according to the underlying symptoms. In order to provide detailed information on how stress production problems can be addressed, the study investigated (1) the minimum amount of change in a single stress marker necessary to achieve significant improvement in stress target identification; and (2) whether stress can be signalled more effectively with a combination of stress markers. Data were sourced from a sentence stress task performed by 10 speakers with ataxic dysarthria and 10 healthy matched control participants. Fifteen utterances perceived as having incorrect stress patterns (no stress, all words stressed or inappropriate word stressed) were selected and digitally manipulated in a stepwise fashion based on typical speaker performance. Manipulations were performed on F0, intensity and duration, either in isolation or in combination with each other. In addition, pitch contours were modified for some utterances. A total of 50 naïve listeners scored which word they perceived as being stressed. Results showed that increases in duration and intensity at levels smaller than produced by the control participants resulted in significant improvements in listener accuracy. The effectiveness of F0 increases depended on the underlying error pattern. Overall intensity showed the most stable effects. Modifications of the pitch contour also resulted in significant improvements, but not to the same degree as amplification. Integration of two or more stress markers did not result in better results than manipulation of individual stress markers, unless they were combined with pitch contour modifications. The results highlight the potential for improvement of stress production in speakers with motor speech disorders. The fact that individual parameter manipulation is as effective as combining them will facilitate the therapeutic process considerably, as will the result that amplification at lower levels than seen in typical speakers is sufficient. The difference in results across utterance sets highlights the need to investigate the underlying error pattern in order to select the most effective compensatory strategy for clients.

An Exploration of Anticipation of Stuttering in Adults

Objective: The experience of stuttering is more complex than its observable behaviors, and in order for a person who stutters (PWS) to truly experience fluent speech the complete experience of stuttering must be removed. This study investigates the effects of various speaking conditions on one aspect of the experience of stutteringanticipation of stuttering. Methods: Ten adult stutterers were asked to read three passages under three reading conditions: normal reading, choral speech, and extremely prolonged speech. Participants made pre anticipatory evaluations assuming normal reading and post anticipatory evaluations assuming reading under a recently performed condition. A Friedman test was utilized to determine the presence of a difference between reading conditions for both dependent variables–stuttering occurrences and changes in anticipation of stuttering. These differences were then examined using a non-parametric Wilcoxon Signed Rank Test. Results: Results from the Friedman tests indicated significant differences between stuttering occurrences and differences in pre and post anticipated events of stuttering at all common significant levels. The non-parametric Wilcoxon Signed Rank Tests revealed significant differences in stuttering occurrences between choral speech and normal reading (p=0.008) and extremely prolonged speech and normal reading (p=0.008). The Wilcoxon also revealed significant differences in anticipation of stuttering between choral speech and normal reading (p=0.005) and extremely prolonged speech and normal reading (p=0.005). Conclusion: This study informs us that for some PWS, anticipation of stuttering events are eliminated while speaking with extremely prolonged speech and/or choral speech. The importance of this finding is the beginning of understanding the difference between perceivably fluent speech (remaining under the control of stuttering) and fluent speech that more closely approximates that of a typical speaker.

The University of Everywhere: Report of the Executive Vice President to the House of Delegates

The University of Everywhere: Report of the Executive Vice President to the House of Delegates

Tongue Dynamics in Childhood Apraxia of Speech: A Case Study

Childhood Apraxia of Speech (CAS) is a subtype of speech sound disorder with unique features that include deficits in speech sound accuracy, prosody, coarticulatory transitions, and consistency on repeated attempts. Articulatory errors are abundant in apraxia such as articulatory groping, perseverative errors, and speech initiation difficulties. There are few studies conducted to see the articulatory pattern of CAS. Present study aimed to analyse the prime articulator, i.e. tongue dynamics in CAS in different place of articulation comparing with typically developing children. One child with CAS and one typically developing age matched control subject were participated in the study. The test material consisted of three meaningful words consisting stops in three different places of articulation including dental /t?/, retroflex/?/ and velar /k/. The instrument Mindray ultrasound 6600 with Articulate Assistant Advanced (AAA) ultrasound module Version 2.14 was used for data collection and analysis. The overall results showed variation across trials in CAS and this disparity between the trails were less in typical speaker’s utterances compared to CAS. The placement of tongue was diverse across subject during the articulation of /k/ with respect to the tongue height and advancement. Variations were more evident in the tongue front and posterior tongue body region compared to anterior tongue body. In the case of retroflex stop /?/ tongue front and posterior tongue body image was not prominent in all repetition of CAS. But in normal speaker all the three divisions were prominent and variations across trails were less. This information can be useful to set the treatment plans for CAS to resolve the articulatory errors.

Automatic Speaker Recognition System in Adverse Conditions — Implication of Noise and Reverberation on System Performance

Speaker recognition has been developed and evolved over the past few decades into a supposedly mature technique. Existing methods typically utilize robust features extracted from clean speech. In real-world applications, especially security and forensics related ones, reliability of recognition becomes crucial, meanwhile limited speech samples and adverse acoustic conditions, most notably noise and reverberation, impose further complications. This paper is presented from a study into the behavior of typical speaker recognition systems in adverse retrieval phases. Following a brief review, a speaker recognition system was implemented using the MSR Identity Toolbox by Microsoft. Validation tests were carried out with clean speech and the speech contaminated by noise and/or reverberation of varying degrees. The image source method was adopted to take into account real acoustic conditions in the spaces. Statistical relationships between recognition accuracy and signal to noise ratios or reverberation times have therefore been established. Results show noise and reverberation can, to different extents, degrade the performance of recognition. Both reverberation time and direct to reverberation ratio can affect recognition accuracy. The findings may be used to estimate the accuracy of speaker recognition and further determine the likelihood a particular speaker.

Coping with Lexical Gaps: Use of Compensatory Strategies by Iranian EFL Students

Learners of English as a non-native anguage unavoidably confront situations in which insufficiency of their linguistic knowledge presents them with difficulties. There are also case in which the absence of some native-language lexical items in the target language is the source of problems the learners have to grapple with. This study focused on the phenomenon known as ˜lexical gap .The aim was to identify and examine the use of compensatory strategies employed by Iranian EFL students in coping with lexical gaps.To this end, number of Iranian undergraduate EFL students were selected and then split into two groups of intermediate-level and advanced-level. Both groups were given two written tasks requiring them to introduce eight Farsi words with no English counterparts to a typical native speaker of English. The analysis of data revealed well as the existence of some surprising differences between the frequencies of strategy use by the two groups.

Improvement in the Acoustic Characteristics of the Micro Speaker by Changing Thickness and Foaming Ratio of the Diaphragm

The typical speaker for cellular phone is composed of magnet, coil, base-film and diaphragm with a structure that laminated a foam sheet and aluminum sheets. A heterogeneous foam sheet which has an uneven elastic modulus was used for a conventional diaphragm, and it reduced an acoustic performance of a conventional speaker. Recently a homogeneous and thin foam sheet was newly developed. The purpose of this study is to improve the acoustic characteristics of the micro speaker by changing thickness and foaming ratio of the newly developed foam sheet. Frequency characteristics and sound pressure level (SPL) were measured by changing foaming ratio of the foam sheet from 2.5 to 5.0 and thickness of the foam sheet from 0.15 mm to 0.30 mm. To clarify the relationship between the design parameter of foam sheet and the frequency characteristics of the micro speaker, natural frequency analysis using finite element method (FEM) was performed. The result of FEM analysis was also compared with the experimental result using a laser doppler velocimetry sensor. It was found that the widest reproducible frequency range was obtained when the foam sheet was used 0.30 mm of thickness and 5.0 of foaming ratio. The reproducible frequency range spread out 15% in comparison with that of a conventional micro speaker.

Synthesized speech for model training in cross-corpus recognition of human emotion

Recognizing speakers in emotional conditions remains a challenging issue, since speaker states such as emotion affect the acoustic parameters used in typical speaker recognition systems. Thus, it is believed that knowledge of the current speaker emotion can improve speaker recognition in real life conditions. Conversely, speech emotion recognition still has to overcome several barriers before it can be employed in realistic situations, as is already the case with speech and speaker recognition. One of these barriers is the lack of suitable training data, both in quantity and quality—especially data that allow recognizers to generalize across application scenarios (‘cross-corpus’ setting). In previous work, we have shown that in principle, the usage of synthesized emotional speech for model training can be beneficial for recognition of human emotions from speech. In this study, we aim at consolidating these first results in a large-scale cross-corpus evaluation on eight of most frequently used human emotional speech corpora, namely ABC, AVIC, DES, EMO-DB, eNTERFACE, SAL, SUSAS and VAM, covering natural, induced and acted emotion as well as a variety of application scenarios and acoustic conditions. Synthesized speech is evaluated standalone as well as in joint training with human speech. Our results show that the usage of synthesized emotional speech in acoustic model training can significantly improve recognition of arousal from human speech in the challenging cross-corpus setting.

Tuning-Robust Initialization Methods for Speaker Diarization

This paper investigates a typical speaker diarization system regarding its robustness against initialization parameter variation and presents a method to reduce manual tuning of these values significantly. The behavior of an agglomerative hierarchical clustering system is studied to determine which initialization parameters impact accuracy most. We show that the accuracy of typical systems is indeed very sensitive to the values chosen for the initialization parameters and factors such as the duration of speech in the recording. We then present a solution that reduces the sensitivity of the initialization values and therefore reduces the need for manual tuning significantly while at the same time increasing the accuracy of the system. For short meetings extracted from the previous (2006, 2007, and 2009) National Institute of Standards and Technology (NIST) Rich Transcription (RT) evaluation data, the decrease of the diarization error rate is up to 50% relative. The approach consists of a novel initialization parameter estimation method for speaker diarization that uses agglomerative clustering with Bayesian information criterion (BIC) and Gaussian mixture models (GMMs) of frame-based cepstral features (MFCCs). The estimation method balances the relationship between the optimal value of the seconds of speech data per Gaussian and the duration of the speech data and is combined with a novel nonuniform initialization method. This approach results in a system that performs better than the current ICSI baseline engine on datasets of the NIST RT evaluations of the years 2006, 2007, and 2009.

Linguistics, Psychology and the Scientific Study of Language

In this paper I address the issue of the subject matter of linguistics. According to the prominent Chomskyan view, linguistics is the study of the language faculty, a component of the mind-brain, and is therefore a branch of cognitive psychology. In his recent book Ignorance of Language Michael Devitt attacks this psychologistic conception of linguistics. I argue that the prominent Chomskyan objections to Devitt's position are not decisive as they stand. However, Devitt's position should ultimately be rejected as there is nothing outside of the mind of a typical speaker that could serve to fix determinate syntactic rules of her language or constitute the supervenience base of her connection to any such rules.

Effects of audio-visual information and mode of speech on listener perceptions of alaryngeal speakers

There is minimal research on listener perceptions of an individual with a laryngectomy (IWL) based on audio-visual information. The aim of this research was to provide preliminary insight into whether listeners have different perceptions of an individual with a laryngectomy based on mode of presentation (audio-only vs. audio-visual) and mode of speech (tracheoesophageal, oesophageal, electrolaryngeal, normal). Thirty-four naïve listeners were randomly presented with a standard reading passage produced by one typical speaker from each mode of speech in both audio-only and audio-visual presentation mode. Listeners used a visual analogue scale (10 cm line) to indicate their perceptions of each speaker's personality. A significant effect for mode of speech was present. There was no significant difference in listener perceptions between mode of presentation using individual ratings. However, principal component analysis showed ratings were more favourable in the audio-visual mode. Results of this study suggest that visual information may only have a minor impact on listener perceptions of a speakers' personality and that mode of speech and degree of speech proficiency may only play a small role in listener perceptions. However, results should be interpreted with caution as results are based on only one speaker per mode of speech.

Stereotyping and Victim Blaming of Individuals with a Laryngectomy

This study examined whether individuals with a laryngectomy are at risk for stereotyping and victim blaming. Undergraduate students (3 men and 41 women) volunteered to listen to a recording of a man who used tracheoesophageal speech or a man who was a typical adult speaker. A written vignette was provided prior to the listening activity to assess whether knowing that the speaker did or did not have a history of smoking affected the listeners' judgments of the speaker. Semantic differential data and qualitative responses indicated that participants did not report victim blaming; however, they did report that users of tracheoesophageal speech were less highly employable than typical adult speakers. Participants also indicated they relied on the vocal characteristics of speakers and used empathy- or experienced-based reactions to make judgments about users of such speech.

SVM-Based Speaker Verification System for Match-on-Card and Its Hardware Implementation

Using biometrics to verify a person's identity has several advantages over the present practice of personal identification numbers (PINs) and passwords. To gain maximum security in a verification system using biometrics, the computation of the verification as well as the storing of the biometric pattern has to take place in a smart card. However, there is an open issue of integrating biometrics into a smart card because of its limited resources (processing power and memory space). In this paper, we propose a speaker verification algorithm using a support vector machine (SVM) with a very few features, and implemented it on a 32-bit smart card. The proposed algorithm can reduce the required memory space by a factor of more than 100 and can be executed in real-time. Also, we propose a hardware design for the algorithm on a field-programmable gate array (FPGA)-based platform. Based on the experimental results, our SVM solution can provide superior performance over typical speaker verification solutions. Furthermore, our FPGA-based solution can achieve a speed-up of 50 times over a software-based solution.

Sound solutions for cellular phones

The ongoing miniaturization of cellular phones during the last years has led to the development of smaller and smaller speakers for these products. Today’s size of a typical speaker for GSM and CDMA phones is of a diameter of 15 and 13 mm, and we are about to see even smaller speakers on the market. In the design and type approval of the receiving part of the cellular phone, the reference is a standardized measuring ear, which seals toward the ear cap plane of the phone. Under these conditions, it is possible to achieve a flat frequency response representing the whole telephone band, even with speakers in small sizes. However, such a tiny speaker in combination with a very small cellular phone having a poor surface of contact to the human ear leads to heavily reduced speech quality, which cannot be reflected in the measurement result. In other words, there is a very poor correlation between the type real life telephone user situation. This presentation gives the guidelines on how to design smaller cellular phones prepared for type approval using small speakers and at the same time improve sound quality.

The art and science of the conference presentation.

I have just returned from a conference in Istanbul. Among my most vivid memories are the panoramic cityscape, the strong Turkish coffee and more than a few extraordinarily poor conference speakers. Of course, some presenters were polished and indeed, inspirational, but a substantial minority of the professors, management consultants and business leaders who spoke bordered on the incompetent. Unfortunately, this is not the first conference that I have attended where this is the case.Audiences at conferences do not expect their colleagues to speak with the wit of David Letterman or the charisma of Oprah Winfrey. Likewise, audiences at conferences may challenge speakers in academic debate but they rarely find that the speaker has nothing to contribute. The most common complaint is that the speaker either lacks the basic skills or the motivation to make a good presentation.This is a disappointing situation in which all stakeholders lose. The ineffective speaker fails to capitalize on the 20 or 30 minutes available in the program to share knowledge born of months or even years of planning, executing and interpreting research. The conference audience loses a chance to be introduced to leading-edge ideas and data. The conference organizers fall into a vicious cycle: ineffective speakers contribute to lowered status for the conference, making it harder for organizers to attract good presenters for the next conference. Both the merit and stature of our conferences will increase if we collectively commit to improving the quality of our presentation style.It would be all too easy for me to end with a simple rallying cry. Let's try harder! Let's plan better! Let's buy PowerPoint for our computers! Such platitudes won't work. Let me instead PRESENT concrete suggestions derived from my experiences as a conference participant who is eager to hear her colleagues' academic insights, but is often frustrated with the presentation process. These suggestions also apply to other contexts in which academics speak about their work, including graduate students making thesis or dissertation presentations.SuggestionsPick Your Audience. Decide who your audience is. Is it the generalist, the specialist in your subarea, or the technical wizard? You cannot convey your information to all groups in the typical 20-minute presentation. Moving back and forth satisfies no one constituency; make your decision and prepare your talk for that particular audience.Respect Time Limits. How would you feel if the evening news ran over time, pre-empting 10 minutes of your favourite television program, because the newscaster was verbose and disorganized? That is exactly how a conference audience feels when a conference speaker goes on and on -- impatient, then frustrated, and finally, hostile. Time management is the single biggest problem for conference speakers.Time management has two aspects. First, you need to figure out how much time you really have to make your presentation. The answer is three or four minutes less than that formally allotted in the program. Consider that it takes time to walk to the front of the room, get out your notes or visual aids, adjust the microphone, and wait for that last straggler to sit down. Likewise, you need a minute or two at the end of the presentation. Think what a good impression it leaves when a speaker finishes calmly, with ample time to collect presentation materials and to thank the audience for its attention.Now consider the other aspect of time management. How much material can you reasonably cover in X minus three minutes? Many speakers dread the thought that they might run out of material so they prepare lots and plan to cut out things as the talk progresses. Other speakers prepare the presentation to cover what they consider to be essential, disregarding the formal time limit entirely. Both these strategies are guaranteed to cause time problems.You can estimate the amount of material you can cover in a presentation by assuming that the typical speaker can cover one overhead every three minutes. …

Speaker normalization of static and dynamic vowel spectral features

Two methods are described for speaker normalizing vowel spectral features: one is a multivariable linear transformation of the features and the other is a polynomial warping of the frequency scale. Both normalization algorithms minimize the mean-square error between the transformed data of each speaker and vowel target values obtained from a "typical speaker." These normalization techniques were evaluated both for formants and a form of cepstral coefficients (DCTCs) as spectral parameters, for both static and dynamic features, and with and without fundamental frequency (F0) as an additional feature. The normalizations were tested with a series of automatic classification experiments for vowels. For all conditions, automatic vowel classification rates increased for speaker-normalized data compared to rates obtained for nonnormalized parameters. Typical classification rates for vowel test data for nonnormalized and normalized features respectively are as follows: static formants--69%/79%; formant trajectories--76%/84%; static DCTCs 75%/84%; DCTC trajectories--84%/91%. The linear transformation methods increased the classification rates slightly more than the polynomial frequency warping. The addition of F0 improved the automatic recognition results for nonnormalized vowel spectral features as much as 5.8%. However, the addition of F0 to speaker-normalized spectral features resulted in much smaller increases in automatic recognition rates.

Long‐term changes in voice characteristics: Implications for speaker identity verification (SIV)

In a typical speaker identity verification (SIV) scenario, a talker first asserts an identity claim via some means other than speech. A reference sample of the claimed individual's speech is then loaded into a pattern matcher and compared against a new sample of the same word or phrase solicited on‐line. Based on the similarity of the two utterances, and a predetermined criterion, a decision is made as to the veracity of the claimed identity. One difficult problem for SIV applications is that the spectral characteristics of a speaker's voice change over time in ways that are poorly understood. Utterances were collected from three speakers at regular intervals over a period of 23 weeks, and were analyzed first with respect to the magnitude of spectral dissimilarity over varying time intervals, and, second, with respect to the effect of these differences on error rates for an SIV decision task. The data show an increase in mean utterance dissimilarity of about 8% over a period of 2 weeks, followed by a much more gradual increase over 6 months. While the absolute change in mean dissimilarity was small, it resulted in a doubling of the error rate for the SIV decision task from about 5% to 10%.In a typical speaker identity verification (SIV) scenario, a talker first asserts an identity claim via some means other than speech. A reference sample of the claimed individual's speech is then loaded into a pattern matcher and compared against a new sample of the same word or phrase solicited on‐line. Based on the similarity of the two utterances, and a predetermined criterion, a decision is made as to the veracity of the claimed identity. One difficult problem for SIV applications is that the spectral characteristics of a speaker's voice change over time in ways that are poorly understood. Utterances were collected from three speakers at regular intervals over a period of 23 weeks, and were analyzed first with respect to the magnitude of spectral dissimilarity over varying time intervals, and, second, with respect to the effect of these differences on error rates for an SIV decision task. The data show an increase in mean utterance dissimilarity of about 8% over a period of 2 weeks, followed by a much...