Automatic speaker recognition represents one of the most transformative frontiers at the intersection of acoustic phonetics, signal processing, and computational intelligence. By translating the idiosyncratic anatomical and behavioral traits of human speech into high-dimensional mathematical representations, this technology empowers machines to authenticate identity with remarkable precision. As voice interfaces increasingly mediate daily interactions, understanding the architectural and cognitive mechanisms underlying automated voice identification becomes essential for technologists, forensic scientists, and cognitive researchers alike.
Automatic Speaker Recognition
1. Concise Definition
Automatic speaker recognition (ASR) is the computational process of automatically validating or determining a person’s identity using the biometric characteristics extracted from their acoustic speech signals. It operates by converting raw pressure waves produced by the human vocal tract into quantifiable digital signatures that reflect both permanent physical anatomy and learned behavioral nuances. Unlike speech recognition, which transcribes what is spoken, speaker recognition verifies who is speaking.
Conceptually, the domain encompasses two primary operational paradigms: speaker verification (also known as voice authentication) and speaker identification. Verification performs a one-to-one (1:1) binary hypothesis test to confirm a claimed identity, while identification executes a one-to-many (1:N) search across an enrolled population to determine the true speaker. These methodologies form the technological backbone of voice biometrics, forensic acoustics, telephonic access control, and seamless person-machine personalization.
2. Etymology & Linguistic Origin
The term “automatic speaker recognition” derives from a synthesis of classical Greek and Latin linguistic roots adapted into modern computational terminology. The adjective “automatic” originates from the Greek automatos (composed of autos, meaning “self”, and matos, meaning “acting, willing, or thinking”), historically describing machines or mechanisms operating via internal triggers without continuous human agency. The noun “speaker” stems from the Old English sprecan (cognate with the Proto-Germanic *sprekanan, “to speak”), signifying an agent producing articulate phonetic utterances.
The constituent “recognition” traces to the Latin recognoscere, formed by the prefix re- (“again”) and cognoscere (“to learn, know, or investigate”). Within acoustic engineering and phonetic sciences during the mid-twentieth century, researchers at institutions such as Bell Labs merged these terms to differentiate machines that perceived semantic language from those dedicated specifically to recognizing individual talker characteristics. Over the subsequent decades, the compound phrase entered standard computer science, engineering, and psycholinguistic nomenclature to designate autonomous biometric identity inference from voice audio.
3. Pronunciation & Grammatical Form
The term is pronounced phonetically in International Phonetic Alphabet (IPA) notation as /ˌɔːtəˈmætɪk ˈspiːkər ˌrɛkəɡˈnɪʃən/ in standard American English, and /ˌɔːtəˈmætɪk ˈspiːkə ˌrɛkəɡˈnɪʃn/ in Received Pronunciation British English.
Grammatically, “automatic speaker recognition” functions as a non-count compound noun phrase. It routinely occupies the position of grammatical subject, direct object, or adjunct noun in specialized discourse (e.g., “Automatic speaker recognition systems exhibit vulnerabilities to non-stationary acoustic noise”). Common variations in academic literature include “automated speaker recognition,” “voiceprint verification,” and “computational voice biometrics.” In technical literature, researchers frequently abbreviate the discipline as SRec or SpkRec to prevent semantic confusion with Automatic Speech Recognition (abbreviated as ASR).
4. Detailed Conceptual Explanation
Automatic speaker recognition operates on the premise that human speech possesses extraordinary inter-individual variability combined with relatively stable intra-individual consistency. The human voice is produced through an intricate interaction between biological hardware and neurological software. The biological hardware consists of the subglottal respiratory system, the vocal folds within the larynx, and the supraglottal vocal tract comprising the pharynx, oral cavity, and nasal passages. The physical geometry, tissue compliance, and dimensional ratios of these anatomical structures impart distinct spectral resonances, known as formants, to the acoustic waveform. Because no two individuals possess identical anatomical dimensions, speech spectra naturally convey individual-specific signatures.
Complementing this physiological baseline are behavioral and dialectal habits acquired across an individual’s lifespan. These behavioral nuances include articulatory agility, idiosyncratic pitch contours, baseline fundamental frequency (F0), glottal pulse shaping, speaking rate, regional phonetic shifts, and prosodic cadences. While anatomical characteristics determine the boundaries of an individual’s acoustic envelope, learned motor execution governs dynamic movement within those boundaries. Automatic speaker recognition architectures are engineered to distill both anatomical and behavioral patterns into compact, robust vector representations while filtering out contextual linguistic variation, environmental noise, and transmission channel distortions.
The fundamental operational pipeline of an automatic speaker recognition system comprises four sequential stages: acoustic signal acquisition, digital pre-processing, feature extraction, and statistical modeling/matching. Pre-processing removes extraneous silence via voice activity detection (VAD), normalizes signal amplitude, and minimizes linear channel mismatch through spectral subtraction or cepstral mean subtraction. Feature extraction subsequently transforms the time-domain waveform into a continuous series of compact representations—historically utilizing Mel-frequency cepstral coefficients (MFCCs) and linear predictive coding (LPC), and more recently relying on deep acoustic embeddings extracted from neural network bottlenecks.
During enrollment, these extracted acoustic features are processed to construct a permanent reference profile of the target individual. In historical systems, this reference profile manifested as a Gaussian Mixture Model (GMM) adapted from a broad Universal Background Model (UBM), or as a low-dimensional identity vector (i-vector). In contemporary architectures, modern deep neural networks map speech samples directly into highly separable latent metric spaces, producing dense fixed-length representations known as x-vectors, ECAPA-TDNN embeddings, or transformer-based latent states. When a probe sample is submitted, the system computes the similarity between the probe embedding and the enrolled model using scoring metrics such as cosine distance or probabilistic linear discriminant analysis (PLDA), determining identity against a predetermined decision threshold.
5. Historical Development
The genesis of automatic speaker recognition emerged from forensic acoustic studies and early telegraphic sound analysis conducted during the mid-twentieth century. In the late 1940s and 1950s, researchers at Bell Laboratories, including Ralph Potter and Lawrence Kersta, explored the concept of acoustic “voiceprints.” Kersta hypothesized that visual inspections of spectrograms displaying frequency, time, and energy could identify human voices with the same deterministic certainty as dermal friction ridges in latent fingerprint analysis. Although subsequent scientific investigations revealed that inter-speaker variability was confounded by intra-speaker emotional and physiological fluctuation, this foundational work stimulated rigorous exploration into mathematical sound modeling.
The transition from manual spectrogram inspection to digital computing accelerated during the 1970s and 1980s. Early algorithmic implementations relied heavily on dynamic time warping (DTW) and filter-bank analyses, largely restricted to text-dependent tasks where speakers uttered identical predefined passwords. During this era, researchers such as Sadaoki Furui pioneered the application of cepstral analysis and orthogonal polynomial representations to capture dynamic spectral trajectories, laying the bedrock for robust acoustic feature extraction.
The 1990s and early 2000s witnessed a major paradigm shift marked by the hegemony of statistical modeling. Douglas Reynolds and Thomas Quatieri formalized the Gaussian Mixture Model-Universal Background Model (GMM-UBM) framework. This breakthrough solved the critical problem of sparse training data by training a massive, speaker-independent acoustic model (the UBM) using expectation-maximization, and subsequently adapting it to individual enrolled speakers through Maximum A Posteriori (MAP) estimation. Shortly thereafter, Patrick Kenny introduced Joint Factor Analysis (JFA) to mathematically disentangle speaker-specific variability from transmission channel and acoustic session variability.
In 2011, Najim Dehak and colleagues streamlined factor analysis into the i-vector paradigm, which condensed both speaker and channel variabilities into a single low-dimensional continuous subspace. Coupled with PLDA classifiers, i-vectors dominated academic benchmarks, including the National Institute of Standards and Technology (NIST) Speaker Recognition Evaluations (SRE), for nearly a decade. By the mid-2010s, the deep learning revolution superseded classical factor analysis. Beginning with the introduction of deep neural network embeddings (x-vectors) by David Snyder and researchers at Johns Hopkins University, contemporary speaker recognition transitioned toward end-to-end architectures, convolution-augmented transformers, and self-supervised foundation models trained on tens of thousands of hours of unconstrained conversational audio.
6. Theoretical Foundations
The conceptual framework underpinning automatic speaker recognition is anchored in the source-filter theory of acoustic speech production, initially formulated by Gunnar Fant in 1960. The source-filter model posits that speech sounds can be mathematically decomposed into two distinct, quasi-independent components: an acoustic excitation source and a linear acoustic resonant filter. The source corresponds to the mechanical vibration of the vocal folds, producing a fundamental harmonic spectrum, while the filter corresponds to the supraglottal vocal tract, which acts as a dynamic acoustic tube that selectively amplifies and attenuates specific harmonic frequencies.
Within the source-filter paradigm, speaker recognition operates across both elements. The physical mass, tension, and glottal geometry of the vocal folds govern vocal timbre, phonation types, and pitch characteristics (the source). Concurrently, the lengths, volumes, and cross-sectional areas of the oral, pharyngeal, and nasal cavities govern formant positions and bandwidths (the filter). Because human adult vocal tract lengths vary significantly—typically averaging 17 cm in males and 14 cm in females, with substantial individual variance within each cohort—the resonant transfer functions generated by the filter provide highly discriminatory physiological markers.
Complementing this physiological acoustics framework is the statistical foundation of pattern recognition and information theory. Speaker recognition can be formulated through Bayesian decision theory, which determines whether an acoustic sample $Y$ originated from an enrolled speaker $S$ or an alternative impostor speaker $\Omega$. Using the log-likelihood ratio (LLR), the classification rule evaluates:
$$\Lambda(Y) = \log \frac{P(Y mid H_0)}{P(Y mid H_1)}$$
Where $H_0$ denotes the null hypothesis that the speech originates from the claimed speaker $S$, and $H_1$ designates the alternative hypothesis that the speech originates from any random speaker within the non-target cohort $\Omega$. When the calculated log-likelihood ratio exceeds an empirical operational threshold $\theta$, the null hypothesis is accepted, confirming identity.
In modern high-dimensional deep representation learning, this theoretical framework is manifested through metric learning paradigms. Deep networks are trained using margin-based classification loss functions, such as Additive Margin Softmax (AM-Softmax) or ArcFace. These loss functions enforce rigorous geometric constraints on the latent unit hypersphere: minimizing intra-class cosine distance among diverse utterances spoken by the same individual, while actively maximizing inter-class angular margins between different individuals. Consequently, the network constructs an invariant identity subspace where acoustic distance correlates directly with human biological and behavioral divergence.
7. Key Components, Types & Dimensions
Automatic speaker recognition architectures can be classified across several operational, technical, and architectural dimensions:
- Speaker Verification vs. Speaker Identification: Verification involves a binary classification (1:1 matching) where an individual claims an identity and the system either accepts or rejects the claim. Identification involves a polychotomous decision (1:N matching) where the system compares an unknown voice against an entire database of enrolled candidates to return the most probable speaker or conclude the speaker is unknown.
- Text-Dependent vs. Text-Independent Systems: In text-dependent systems, the user must pronounce an exact predefined passkey or prompted sequence of words (e.g., “My voice is my password”). In text-independent systems, the model identifies the speaker regardless of vocabulary, topic, or conversational structure, relying purely on text-invariant acoustic properties.
- Short-Term Spectral vs. Long-Term Suprasegmental Features: Short-term spectral features capture instantaneous vocal-tract geometry across 20-30 millisecond frames (e.g., MFCCs, filter-bank energies). Long-term suprasegmental features capture prosody, speaking rate, rhythm, pitch variance, and dynamic phonetic transitions over entire phrases or conversations.
- Front-End Acoustic Feature Extractors: The computational module responsible for reading digital audio signals, stripping acoustic silence, normalizing background noise, and transforming raw waveforms into structured multidimensional arrays (e.g., short-time Fourier transforms or learned raw-waveform convolutions).
- Back-End Embedding Extractors: Deep learning architectures optimized for mapping variable-length speech sequences into compact, fixed-dimensional speaker vectors (e.g., Time-Delay Neural Networks [TDNN], ResNet-based architectures, ECAPA-TDNN, and self-supervised wav2vec or HuBERT representations).
- Scoring and Decision Backend: The statistical or geometric metric engine that compares the target speaker embedding with the reference enrollment model (e.g., Cosine Similarity Scoring, Mahalanobis Distance, or Probabilistic Linear Discriminant Analysis [PLDA] with score normalization algorithms such as Z-norm and S-norm).
8. Examples & Illustrative Cases
To conceptualize the execution of automatic speaker recognition across distinct functional contexts, consider the following real-world operational scenarios:
Case 1: Frictionless Financial Voice Biometrics. A commercial banking client places a telephone call to an automated customer support terminal. Rather than requiring the client to recall complex alphanumeric passwords, security questions, or mother’s maiden names, the system initiates a text-independent speaker verification pipeline. As the client naturally explains the reason for their call, the front-end processor extracts speech segments, filters out telephonic transmission channel artifacts (such as 8 kHz band-limiting and adaptive multi-rate codec distortions), and projects the acoustic frames through a trained neural embedding network. Within three seconds of natural dialogue, the resulting vector achieves a high log-likelihood ratio against the customer’s stored reference profile, quietly authenticating the user without interrupting conversational flow.
Case 2: Forensic Speaker Identification in Criminal Investigations. During a high-profile extortion investigation, law enforcement agencies recover a series of threatening, anonymous voicemail recordings containing heavy background noise and cellular compression. Forensic acousticians deploy a calibrated text-independent speaker identification system. The system segments the audio to isolate intelligible speech, applies voice activity detection to remove environmental reverberation, and extracts speaker vectors. These vectors are cross-matched against a database comprising several hundred known suspects using a Bayesian Likelihood Ratio framework. The output provides the court not with an absolute binary declaration of guilt, but with a statistically calibrated strength of evidence reflecting the relative likelihood of observing the acoustic data under the prosecution versus defense hypotheses.
Case 3: In-Cabin Multi-Speaker Personalization. A modern smart vehicle integrates far-field microphone arrays inside the passenger cabin. When an occupant issues the generic command, “Navigate to my office,” the integrated speaker recognition engine determines whether the acoustic vector corresponds to the primary driver, the co-driver, or a rear-seat passenger. Despite acoustic interference from road vibration, air conditioning, and rolling tyre noise, the system matches the embedding to the primary driver’s profile, automatically populating their personal calendar destinations, adjusting their preferred seat positions, and calibrating individual climate controls.
9. Measurement & Assessment
Assessing the efficacy, robustness, and reliability of automatic speaker recognition architectures demands rigorous empirical metrics grounded in detection theory. Because verification operates as a binary hypothesis test, performance is primarily defined by the distribution of two complementary error modes: False Acceptance (Type I Error) and False Rejection (Type II Error).
- False Acceptance Rate (FAR): The proportion of verification attempts by unauthorized impostors that are mistakenly accepted by the system as authentic matches. In identity-critical security systems, the FAR must be minimized to near zero.
- False Rejection Rate (FRR): The proportion of verification attempts by genuine, enrolled target speakers that are erroneously rejected by the system. A high FRR results in degraded user experience, user frustration, and excessive fallback friction.
- Equal Error Rate (EER): The operating point where the False Acceptance Rate precisely equals the False Rejection Rate. The EER serves as the standard, benchmark metric across competitive evaluations to quantify overall system discrimination ability independent of a specific operational threshold.
- Detection Error Tradeoff (DET) Curves: A specialized logarithmic graphical plot that displays the non-linear trade-off between the False Rejection Rate (y-axis) and the False Acceptance Rate (x-axis) across all possible decision thresholds. The DET curve allows engineers to select an operational threshold matched precisely to their security versus usability preferences.
- Minimum Detection Cost Function (minDCF): A standardized metric introduced by the NIST Speaker Recognition Evaluations that weights false acceptances and false rejections according to their economic or operational consequences, incorporating prior probabilities of target versus impostor trials:$$ext{DCF} = C_{ ext{Miss}} \cdot P( ext{Target}) \cdot ext{FRR} + C_{ ext{FA}} \cdot (1 – P( ext{Target})) \cdot ext{FAR}$$Where $C_{ ext{Miss}}$ and $C_{ ext{FA}}$ represent relative assigned costs for misses and false alarms.
10. Applications & Practical Significance
The practical applications of automatic speaker recognition span multiple commercial, security, administrative, and clinical sectors. In the financial services sector, voice biometrics has largely replaced legacy knowledge-based authentication methods across enterprise call centers. By passively authenticating customers based on voiceprints, organizations dramatically truncate average handle times (AHT), minimize credential-stuffing vulnerability, and elevate consumer trust.
In telecommunications and consumer electronics, speaker recognition serves as an indispensable enabling layer for voice-first operating systems and smart home hubs (e.g., Apple Siri, Google Assistant, Amazon Alexa). By implementing lightweight on-device speaker verification, these systems distinguish between family members, personalize music playlists and calendar notifications, and safeguard voice-activated financial transactions against unauthorized family members or ambient television playback.
In legal and forensic domains, forensic voice comparison provides judicial bodies with probabilistic assessments of evidentiary recordings. Standardized speaker recognition models assist forensic phoneticians in evaluating intercepted communications, wiretaps, kidnapping ransom demands, and fraudulent claims. Furthermore, in correctional facilities, automated voice tracking systems constantly monitor inmate phone networks to detect illicit cross-inmate communication, synthetic impersonation, and fraudulent call forwarding.
In healthcare and clinical neuropsychology, variations in speaker recognition embeddings are increasingly repurposed as sensitive digital biomarkers. Pathological alterations in vocal fold vibration, prosodic cadence, and dynamic spectral control can indicate early-stage neurodegenerative disorders. Machine learning frameworks fine-tuned on longitudinal voice embeddings assist clinicians in detecting the subtle onset of Parkinson’s disease, Amyotrophic Lateral Sclerosis (ALS), clinical depression, and mild cognitive impairment well before overt behavioral symptoms manifest.
11. Research & Empirical Evidence
Decades of empirical evaluations coordinated by the National Institute of Standards and Technology have driven systematic algorithmic progress in speaker recognition. Benchmark datasets such as VoxCeleb 1 & 2 (Nagrani et al., 2017; Chung et al., 2018) have catalyzed the transition from controlled acoustic environments to real-world “in-the-wild” conditions characterized by background audio clutter, laughter, interruptions, and varying room reverberations.
Key empirical milestones include the development of the x-vector architecture by Snyder et al. (2018), which demonstrated that deep Time-Delay Neural Networks trained on massive sets of multi-speaker corpora substantially outperformed traditional i-vector and PLDA baselines. Their findings established that temporal context across wide acoustic frames enabled neural networks to capture subtle invariant vocal traits that static frame-level models overlooked.
Further empirical breakthroughs emerged with the integration of Squeeze-and-Excitation mechanisms and multi-scale aggregation in the ECAPA-TDNN model introduced by Desplanques et al. (2020). ECAPA-TDNN demonstrated that channel attention mechanisms, which dynamically weight specific feature maps according to their speaker-discriminative value, yielded significant reductions in Equal Error Rates across both text-dependent and text-independent challenges, achieving EERs below 1.0% on standardized VoxCeleb benchmarks.
Contemporary empirical research explores self-supervised learning representations. Groundbreaking work by researchers examining self-supervised speech foundation models—such as wav2vec 2.0 (Baevski et al., 2020) and HuBERT (Hsu et al., 2021)—reveals that massive neural representations pre-trained on tens of thousands of hours of unlabeled audio acquire deep phonetic and talker representations that can be fine-tuned with minimal enrollment data, retaining high accuracy even under extreme temporal constraints (such as speech segments lasting less than two seconds).
12. Cultural & Cross-Cultural Considerations
The cross-linguistic and cross-cultural generalization of automatic speaker recognition architectures represents a vital dimension of systemic equity and reliability. Historically, acoustic models trained predominantly on monolingual datasets (typically Standard American English or European languages) exhibited pronounced performance degradation when applied to tonal languages, accented populations, or linguistically diverse cohorts.
In tonal languages such as Mandarin Chinese, Vietnamese, and Yoruba, variations in pitch (fundamental frequency, F0) convey lexical and semantic meaning rather than purely emotional or individual behavioral idiosyncrasies. When classical speaker recognition systems placed excessive statistical weight on mean fundamental frequency as a speaker-discriminant feature, their false-rejection rates surged among tonal speakers due to lexical pitch fluctuations. Modern architectures must therefore decouple phonetic pitch modulation from physiological glottal characteristics to maintain consistent verification performance across tonal and non-tonal speech.
Furthermore, socio-cultural conventions governing vocal delivery—such as conversational pauses, honorific registers, prosodic deference, vocal fry, and expressive pitch variation—vary substantially across cultures. Speakers navigating high-context communication environments may alter their acoustic registers dramatically between formal institutional contexts and informal family settings. If a user enrolls their voiceprint in a quiet, formal register and later attempts verification under casual or emotionally heightened circumstances, severe intra-speaker mismatch can occur. Addressing these cross-cultural dynamics demands expansive international corpora and culturally calibrated data augmentation techniques.
13. Criticisms, Debates & Limitations
Despite its technical sophistication, automatic speaker recognition confronts significant scientific debates, ethical criticisms, and vulnerabilities:
- Vulnerability to Presentation and Spoofing Attacks: The proliferation of high-fidelity voice cloning, text-to-speech (TTS) synthesis, and voice conversion (VC) technologies poses an existential challenge to voice biometrics. Malicious actors can generate synthetic voice replicas using brief audio samples scraped from social media. Without robust presentation attack detection (PAD) and anti-spoofing countermeasures, conventional verification models are susceptible to exploitation.
- Adversarial Perturbations: Research demonstrates that speaker recognition models are vulnerable to imperceptible acoustic adversarial noise. By adding mathematically calibrated, human-inaudible noise patterns to an impostor’s audio signal, attackers can cause state-of-the-art neural networks to misclassify unauthorized speakers as enrolled targets.
- Acoustic and Environmental Brittleness: While performance is near-perfect in quiet, controlled laboratory environments, it degrades under acoustic mismatch. Incompatible microphone hardware, dynamic room reverberation, non-stationary industrial noise, and telephonic compression codecs degrade performance, yielding unacceptable spikes in False Rejection Rates.
- Aging and Health Variability: The human voice changes over time. Longitudinal anatomical changes—including vocal fold thinning, ossification of laryngeal cartilages, and respiratory capacity decline—naturally shift a person’s acoustic profile over decades. Similarly, transient physiological conditions such as laryngitis, rhinitis, dental surgery, psychological stress, and clinical fatigue can alter vocal tract transfer functions, causing authentication failures.
- Surveillance and Privacy Infringements: The non-cooperative, remote nature of voice capture raises profound civil liberty concerns. Unlike latent fingerprints or retinal scans, which require active user compliance or close physical proximity, acoustic signals can be intercepted surreptitiously across long distances or tapped from passive smart microphones, enabling covert mass surveillance without informed consent.
14. Related Terms & Distinctions
To ensure academic clarity, automatic speaker recognition must be carefully distinguished from related concepts across acoustic processing, phonetics, and machine learning:
- Automatic Speech Recognition (ASR): While speaker recognition determines who is speaking, speech recognition determines what is spoken. Speech recognition explicitly normalizes away individual-specific variations in acoustic timbre and accent to transcribe linguistic phonemes into textual characters.
- Speaker Diarization: Often summarized as answering “who spoke when,” speaker diarization is the computational task of partitioning an unconstrained audio stream containing multiple interacting speakers into homogeneous segments grouped by individual identity, without requiring prior enrollment of the participants.
- Voiceprint vs. Fingerprint: The term “voiceprint” is an analogical misnomer popularized by early media. Unlike a biological friction ridge fingerprint, which remains structurally invariant throughout an individual’s lifetime, an acoustic voice signature is dynamic, mutable, and heavily influenced by emotional arousal, respiratory health, aging, and ambient acoustics.
- Speech Synthesis (Text-to-Speech / TTS): The inverse process of speech recognition; computational algorithms that translate written text into synthetic artificial speech. In modern security contexts, speech synthesis is frequently evaluated as an attack vector against speaker recognition systems.
- Acoustic Language Identification (LID): The autonomous task of determining the specific spoken language being communicated within an audio recording, independent of the personal identity of the individual speaker or the contextual semantic content.
15. Summary / Key Takeaways
Automatic speaker recognition is an advanced biometric methodology that harnesses the unique anatomical geometry of the vocal tract alongside acquired behavioral speech patterns to autonomously authenticate or identify human beings. By converting dynamic acoustic pressure waves into compact, discriminative latent vector representations (embeddings), speaker recognition systems facilitate seamless, hands-free authentication across banking, mobile computing, forensics, and consumer devices.
While contemporary architectures powered by deep neural networks achieve high accuracy on standardized benchmarks, the domain faces ongoing challenges from acoustic mismatch, physical aging, voice cloning, and synthetic adversarial attacks. The evolution of modern speaker recognition centers on building unified self-supervised architectures, refining anti-spoofing defenses, and ensuring cross-cultural equity across all global populations.
References
- Baevski, A., Zhou, Y., Mohamed, A., & Auli, M. (2020). wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33, 12449–12460. https://proceedings.neurips.cc/paper/2020/hash/92d1e1eb1cd6f9f404e47d626aa82d96-Abstract.html
- Chung, J. S., Nagrani, A., & Zisserman, A. (2018). VoxCeleb2: Deep speaker recognition. Interspeech 2018, 1086–1090. https://doi.org/10.21437/Interspeech.2018-1929
- Dehak, N., Kenny, P. J., Dehak, R., Dumouchel, P., & Ouellet, P. (2011). Front-end factor analysis for speaker verification. IEEE Transactions on Audio, Speech, and Language Processing, 19(4), 788–798. https://doi.org/10.1109/TASL.2010.2064307
- Desplanques, B., Thienpondt, J., & Demuynck, K. (2020). ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification. Interspeech 2020, 3830–3834. https://doi.org/10.21437/Interspeech.2020-2650
- Fant, G. (1960). Acoustic Theory of Speech Production. Mouton & Co.
- Hsu, W. N., Bolte, B., Tsai, Y. H. H., Lakhotia, K., Salakhutdinov, R., & Mohamed, A. (2021). HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29, 3451–3460. https://doi.org/10.1109/TASLP.2021.3122291
- Nagrani, A., Chung, J. S., & Zisserman, A. (2017). VoxCeleb: A large-scale speaker identification dataset. Interspeech 2017, 2616–2620. https://doi.org/10.21437/Interspeech.2017-950
- Reynolds, D. A., Quatieri, T. F., & Dunn, R. B. (2000). Speaker verification using adapted Gaussian mixture models. Digital Signal Processing, 10(1–3), 19–41. https://doi.org/10.1006/dspr.1999.0361
- Snyder, D., Garcia-Romero, D., Sell, G., Povey, D., & Khudanpur, S. (2018). X-vectors: Robust DNN embeddings for speaker recognition. 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 5329–5333. https://doi.org/10.1109/ICASSP.2018.8461375