Indonesian J our nal of Electrical Engineering and Computer Science V ol. 42, No. 3, June 2026, pp. 902 ∼ 912 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v42.i3.pp902-912 ❒ 902 V oice portraits: b uilding faces thr ough v oice analysis Anandhu T . G., J ohn K. J oseph, Na vneeth Krishnan J ., Richu Shib u, Elizabeth Isaac Department of Computer Science and Engineering, Mar Athanasius Colle ge of Engineering, Ernakulam, India Article Inf o Article history: Recei v ed Jul 16, 2025 Re vised Feb 22, 2026 Accepted Mar 4, 2026 K eyw ords: Biometric identication Cross-modal learning Deep learning F acial reconstruction Generati v e adv ersarial netw orks Speech analysis V oice-to-f ace generation ABSTRA CT Generation of a person’ s appearance from their v oice alone i s an intriguing challenge. The proposed frame w ork centers on recreating a person’ s f acial im- age based solely on a short audio recording of that person speaking. Using a deep neural netw ork trained on millions of Y ouT ube recordi ngs where f aces and v oices appear together , the system learns v oice-f ace relationships, enabling it to generate images that capture ph ysical traits such as age, gender , and eth- nicity . Operating in a self-supervised manner , this method tak es adv antage of the pairing of f aces and v oices in online videos, eliminating the need for e x- plicit property modeling. T he model achie v ed a classication accurac y of (95%) for gender , (83%) for age, and (65%) for race prediction from v oice inputs, demonstrating an e xceptional performance in demographic trait identication. The generated images are e v aluated ag ainst real photographs of the speak ers, assessing ho w closely these reconstructions resemble actual appearance. This frame w ork has practical applications in forensic analysis, security systems, and pri v ac y-conscious biometric identication, of fering a non-in v asi v e alternati v e to traditional f acial recognition methods. This is an open access article under the CC BY -SA license . Corresponding A uthor: John K. Joseph Department of Computer Science and Engineering, Mar Athanasius Colle ge of Engineering K othamang alam, Ernakulam, K erala, India Email: johnk.josef@gmail.com 1. INTR ODUCTION Reconstructing f acial features from v oice data is an emer ging area in biometric identication, of fering ne w possibilities for pri v ac y-conscious identication and forensic applications. While traditional f acial recog- nition methods rely hea vily on visual data, v oice-based f acial reconstruction of fers an inno v ati v e and sophisti- cated alternati v e that uses speech characteristics to infer f acial traits. The proposed V oice-to-F ace Generation Frame w ork utilizes adv anced machine learning models, such as V GG netw orks, to generate approximate f acial representations from short v oice recordings. This approach is based on the established connections between a person’ s v oice and ph ysical appearance, where features such as age, gender , and f acial structure subtly inuence v ocal traits lik e pitch, tone, and enunciation [1]. By capturing these relationships, the system aims not to create e xact f aces, b ut to highlight prominent f acial features that align with the unique v ocal characteristics of each indi vidual. T o achie v e this, the audio encoder neural netw ork processes a detailed spectrogram of the input v oice se gment and transforms it into a high-dimensional feature v ector . This v ector , e xtracted from the penultimate layer of a pre-trained f acial recognition model, encodes essential f acial information into a 1024-dimensional feature space. Using the A VSpeech dataset, the model is trained in a self-supervised manner , eliminating the need for e xtensi v e human annotations and enhancing scalability and pri v ac y [2]. J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 903 The relationship between human v oic es and f acial characteristics has been e xtensi v ely studied in cog- niti v e science and biometrics. Research demonstrates that humans can associate unseen f aces with v oices at rates s ignicantly higher than chance, suggesting inherent correlations between v ocal and visual features [1]. These correlations stem from ph ysiological f actors: v ocal tract dimensions, f acial bone structure, and soft tissue characteri stics all inuence both v oice production and f acial appearance. The proposed frame w ork le v er - ages these natural associations through deep learning, enabling automated e xtraction of f acial traits from v oice signals. Recent adv ances in generati v e adv ersarial netw orks (GANs) ha v e sho wn promising results in cross- modal generation tasks. The W a v2Pix frame w ork demonstrated that GANs can ef fecti v ely map speech features to f acial images, achie ving reasonable accurac y in speak er identity matching [2]. Similarly , the Disjoint Map- ping Netw ork (DIMNet) introduced a no v el approach for cross-modal biometric matching, mapping f aces and v oices to a shared co v ariate space [3]. These de v elopments pro vide a strong foundation for the current w ork, which e xtends these concepts to generate detailed f acial reconstructions with enhanced accurac y . The inno v ation of this w ork lies in its combination of w asserstein GAN with gradient penalty (WGAN- GP) for stable training, self-supervised learning from lar ge-scale video datasets, and comprehensi v e e v aluation across multiple demographic attrib utes. Unlik e pre vious approaches that focus primarily on identity match- ing, this frame w ork generates complete f acial images that capture age, gender , and race characteristics with e xceptional precision. This capability opens ne w a v enues for applications in security , forensics, and pri v ac y- preserving biometric systems. 2. RELA TED W ORKS T e xt-based human f ace generation has progressed signicantly , focusing on bridging the g ap between te xt descriptions and visual representations [4]. Researchers ha v e introduced a local-to-global frame w ork em- plo ying graph neural netw orks to model f acial geometry and appearance. These netw orks e xploit the interde- pendencies between f acial components, recognizing that geometry and appearance traits are interrelated and follo w specic distrib utions. This frame w ork generates high-quality , attrib ute-conditioned f acial images from te xtual descriptions, addressing the comple xity of mapping linguistic input to visual output. Extensi v e e xperi- ments v alidate the method’ s ef fecti v eness and usability o v er pre vious approaches. The relationship between human f aces and v oices has been studied e xtensi v ely [1], with ndings demonstrating that humans can associate unseen f aces wi th v oices at rates signicantly higher than chance. Researchers de v eloped a dataset annotated with demographic and audio visual information to computationally model o v erlapping features between f aces and v oices. The results highlight the ef cac y of cross-modal repre- sentations in identifying matching f aces and v oices, adv ancing the understanding of audio visual inte gration. Reconstructing human f aces from ra w speech input has been e xplored using GANs [2], [5]. This approach compares speak er identities in training datasets with generated f acial images, using cross-modal matching for performance e v aluation. The ndings re v eal that the model produces f acial images that align with speak ers’ biometric traits with accurac y f ar e xceeding chance. V oice proling for f ace reconstruction has been addressed through a GAN-based frame w ork that maps speak er identities t o f acial features. This method achie v es accurate f ace generation, v alidated through cross-modal matching techniques, and demonstrates the potential of le v eraging v oice data for biometric applications. Disjoint Mapping Netw ork (DIMNet) [3] has been proposed for cross-modal biometric matching be- tween f aces and v oices. Unlik e traditional approaches, DIMNet maps each modality to a shared co v ariate space to create unied representations. Empirical results sho w that DIMNet outperforms state-of-the-art tech- niques while requiring fe wer data and computational resources, pro viding a promising solution for cross-modal biometric tasks. The speech fusion to f ace (SF2F) [6] frame w ork has been introduced to address challenges in gener - ating f acial images from speech features. This approach impro v es the connection between image generation models and speech domains, resulting in enhanced image quality and feature alignment. Comparati v e studies demonstrate that SF2F achie v es better performance than e xisting methods, making it a rob ust frame w ork for speech-to-f ace generation. Matching speak er audio snippets to f acial images has been studied using con v olutional neural net- w orks (CNNs) [3]. Researchers e v aluated binary and multi-w ay matching tasks using publicly a v ailable datasets, establishing human performance as a baseline. The ndings re v eal that CNNs can surpass human V oice portr aits: b uilding faces thr ough voice analysis (Anandhu T G) Evaluation Warning : The document was created with Spire.PDF for Python.
904 ❒ ISSN: 2502-4752 accurac y in certain scenarios, particularly in dynamic tes ting using video data. Cross-modal identication in speech perception has re v ealed strong connections between auditory and visual modalities. Studies sho w that participants can reliably match unkno wn f aces to v oice s using dynamic stimuli and delayed matching tasks. These ndings underscore the signicance of dynamic information in cross-modal matching, demonstrating that identity-specic cues are shared across modalities [1]. Recent adv ances in deep learning ha v e enabled more sophisticated cross-modal generation approaches. Generati v e models such as v ariational autoencoders (V AEs) and GANs ha v e sho wn remarkable success in syn- thesizing realistic images from v arious input modalities [7]-[25]. The WGAN-GP has been particularly ef fec- ti v e in stabilizing training and impro ving generation quality [12], [26], [27]. These de v elopments pro vide the technical foundation for the current w ork’ s approach to v oice-to-f ace generation. Man y of the re vie wed papers f ace challenges such as a lack of rob ust generaliz ation, where models perform well on specic datasets b ut struggle with unseen real-w orld scenarios. The reliance on e xtensi v e labeled data, which is often scarce or costly to obtain, limits the scalability of these approaches. Additionally , methods focusing on generating f acial images from te xt or v oice data often suf fer from modality mismatch, where linguistic, auditory , and visual cues do not align seamlessly , reducing the delity of the generated im- ages. Ov ertting is another prominent issue, as models tend to memorize training data b ut f ail to generalize to no v el combinations of attrib utes or inputs. Furthermore, man y approaches lack c o m putational ef cienc y , making them unsuitable for real-time or lar ge-scale applications. Our proposed frame w ork aims to address these dra wbacks by emplo ying adv anced cross-modal learning techniques, le v eraging shared latent spaces to align modalities ef fecti v ely , and incorporat ing ef cient neural architectures to impro v e scalability and real-time performance. 3. METHOD The proposed system follo ws a structured multimodal learning frame w ork that maps audio feat u r es to f acial images using a WGAN-GP . This implementation replaces traditional metric learning approaches, such as triplet loss, with a generati v e model that synthesizes realistic f acial images directly from audio inputs. The pipeline consists of multiple stages, including data preprocessing, feature e xtraction, generati v e modeling, and inference. The o v erall architecture of the proposed frame w ork is il lustrated in Figure 1, which sho ws the com- plete pipeline from audio input to f acial image generation. The syste m be gins with audio preprocessing to e xtract Mel spectrograms, which are then encoded into feature v ectors. These features serv e as conditional inputs to a generati v e model that synthesizes f acial images through a series of upsampling operations. The generated images are e v aluated by a critic netw ork that ensures both realism and alignment with the input v oice characteristics. Figure 1. Architecture diagram of the proposed v oice-to-f ace generation frame w ork, sho wing the complete pipeline from audio input through feature e xtraction to f acial image synthesis Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 902–912 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 905 3.1. Data do wnload and pr epr ocessing T o b uild a rob ust dataset, the system rst do wnloads audio and video samples from the A VSpeech dataset. A CSV le containing video IDs and timestamps is parsed to automate the do wnloading process using yt-dlp. Both videos and their corresponding audio les are stored in structured directories to ensure or g anized access. A logging system is inte grated to track processing steps and a v oid redundant do wnloads. The video processing module e xtracts meaningful visual data from the do wnloaded videos. V ideos are rst resampled to maintain uniform frame rates and durations using f fmpe g. Once resampled, frames are e xtracted, focusing on the rst six frames of each video to capture a representati v e f acial appearance. F ace detection is performed using OpenCV’ s deep learning-based model, ensuring that only high-condence (0.5) f aces are retained. Detected f aces are then cropped and resized to RGB images, preparing them for subsequent deep learning models. Simultaneously , audio processing e xtracts speech features that are essential for learning v oice-f ac e correlations. The system con v erts non-.w a v les using p ydub, ensuring uniformity across the dataset. Mel- frequenc y cepstral coef cients (MFCCs) and Mel spectrograms are computed using librosa, pro viding a com- pact yet informati v e representation of speech characteristics. The spectrograms are then resized to RGB pix els, matching the image size used in the generati v e model. 3.2. F eatur e extraction and generati v e modeling Unlik e con v entional embedding-based retrie v al systems, the proposed approach emplo ys a deep gen- erati v e model to synthesize realist ic f aces from audio representations. The fra me w ork consists of three primary components: Audio Encoder . The audio encoder is a con v olutional neural netw ork that transforms RGB Mel spec- trograms into 512-dimensional embeddings [14], [15]. The encoder consists of multiple con v olutional layers that progressi v ely do wnsample the input through batch normalization and ReLU acti v ations. The e xtracted fea- ture v ectors pro vide a compressed representation of the speech signal, preserving essential information about the speak er’ s identity . Generator (F ace synthesis model). The generator tak es the audio embedding and a random noise v ector as inputs and synthesizes a realistic RGB grayscale f acial image. The architecture consists of fully connected layers follo wed by transposed con v olutional layers that progressi v ely upsample the latent space into an image. Batch normalization and T anh acti v ation are applied to stabilize the training process [28]. The generator ef fecti v ely learns to map speech representations to corresponding f ace images, capturing speak er - specic visual attrib utes. Critic (WGAN discriminator). Instead of a traditional discriminator used in standard GANs, the sys- tem emplo ys a w asserstein critic with spectral normalization to e v aluate the authenticity of generated images [9], [27]. The critic consists of multiple con v olutional layers that do wnsample the input f ace images, e xtracting deep feature representations. Additionally , it incorporates a conditional input, taking both the generated image and the corresponding audio embedding to ensure the synthesized f ace maintains alignment with the speak er’ s identity . 3.3. T raining with WGAN-GP The system is trained using the WGAN-GP , which stabilizes training and mitig ates mode collapse issues commonly encountered in GAN-based models. The training process in v olv es the follo wing steps: 3.3.1. A udio encoding The input Mel spectrograms are passed through the audio encoder , generating a 512-dim ensional feature v ector that represents the speak er’ s v oice. 3.3.2. Critic update (Discriminator step) The critic e v aluates both real and generated images to compute a w asserstein loss. A gradient penalty term is added to enforce the Lipschitz constraint, stabilizing the training process. 3.3.3. Generator update The generator synthesizes f aces using the audio embeddings and random noise as input. The critic’ s output is used as a loss signal, encouraging the generator to produce more realistic images that align with speak er identities. V oice portr aits: b uilding faces thr ough voice analysis (Anandhu T G) Evaluation Warning : The document was created with Spire.PDF for Python.
906 ❒ ISSN: 2502-4752 The training follo ws a 5:1 ratio, where the critic is updated v e times for e v ery generator update. This helps maintain a balanced learning process, pre v enting the generator from o v erpo wering the critic too quickly . The optimizer used is Adam [29], with a learning rate of 1 × 10 − 4 and β = (0 . 0 , 0 . 9) . Gradient penalty for stability . T o enforce the Lipschitz cons traint, a gradient penalty (GP) is computed by interpolating between real and generated images. The penalty term ensures that the gradients ha v e a unit norm, pre v enting instability and mode collapse. The GP loss is dened as: GP = λ · ( ∥∇ ˆ x D ( ˆ x, a ) ∥ 2 − 1) 2 (1) where D ( ˆ x, a ) represents the critic’ s output for an interpolated image ˆ x and its corresponding audio embedding a . The penalty weight λ is set to 10.0. 3.4. F ace generation fr om audio Once the model is trained, it can generate a realistic f ace image from a gi v en audio input. The inference process follo ws these steps: A test audio le is loaded and con v erted into a RGB Mel spectrogram. The audio encoder e xtracts a 512-dimensional feature v ector from the spect rogram. A random noise v ector is sampled and concatenated with the audio embedding. The generator synthesizes a corresponding RGB grayscale f ace image. The output image is displayed alongside the ground truth f ace to visually assess the quality of the generated results. 3.5. Ev aluation and visualization The system is e v aluated both qualitati v ely and quantitati v ely . The qualitati v e e v aluation is performed by visually inspecting the generated images and comparing them with ground truth f aces. The quantitati v e e v aluation is done using ince p t ion scores and Fr ´ echet inception distance (FID) to measure the realism and di v ersity of the generated f aces. T o f acilitate i nterpretation, the system includes a GUI-based visualization tool, where users can input an audio le and observ e the generated f ace. The GUI enables easy testing of the model and interacti v e e xploration of dif ferent speak er identities. 4. RESUL TS AND DISCUSSION 4.1. Classication perf ormance The classication performance of the proposed model for predicting gender , age, and race from v oice inputs w as assessed using confusion matrices. The detailed results for each classication task are presented belo w . T able 1 presents the gender classication results, sho wing the confusion matrix for male and female v oice classication. W ith a sample of 100 indi viduals, the model achie v ed a 95% accurac y rate. T able 1. Gender classication results (Accurac y: 95%) Actual \ Predicted 0 (Male) 1 (Female) 0 (Male) 55 3 1 (Female) 2 40 T able 2 sho ws the age classication performance. The model achie v ed an 83% accurac y rate, distin- guishing between younger and older indi viduals. T able 2. Age classication results (Accurac y: 83%) Actual \ Predicted 0 (Y oung) 1 (Older) 0 (Y oung) 41 9 1 (Older) 8 42 T able 3 presents the race classication results. The model achie v ed a 65% accurac y rate in ident ifying racial characteristics from v oice inputs. T able 3. Race classication results (Accurac y: 65%) Actual \ Predicted 0 (Race A) 1 (Race B) 0 (Race A) 28 15 1 (Race B) 20 37 Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 902–912 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 907 4.2. V isualization of classication r esults The confusion matrices for gender , age, and race classication are visualized in Figure 2, pro viding a comprehensi v e vie w of the model’ s classication performance. The confusion matrices indicate that while gender classication remains highly accurate, age and race classication sho w greater v ariance, highlighting areas for future model impro v ement. 55 3 2 40 Actual Predicted 0 1 0 1 (a) Gender (95%) 41 9 8 42 Actual Predicted 0 1 0 1 (b) Age (83%) 28 15 20 37 Actual Predicted 0 1 0 1 (c) Race (65%) Figure 2. Confusion matrices for (a) gender , (b) age, and (c) race classication from v oice inputs, updated to reect real-w orld v ariances in accurac y 4.3. Generated face quality Figure 3 presents sample generated f acial images reconstructed from v oice inputs. Figure 4 s ho ws the corresponding ground truth f acial images used for comparison. The visual comparison demonstrates the model’ s ability to capture k e y f acial features and demographic characteristics from v oice inputs. Figure 3. Sample generated f acial images reconstructed from v oice inputs, demonstrating the model’ s capability to reconstruct f acial features from audio characteristics V oice portr aits: b uilding faces thr ough voice analysis (Anandhu T G) Evaluation Warning : The document was created with Spire.PDF for Python.
908 ❒ ISSN: 2502-4752 Figure 4. Corresponding ground truth f acial images used for comparison with the generated f acial images sho wn in Figure 3 4.4. Discussion of ndings The results demonstrate that the proposed frame w ork achie v es strong performance in demographic trait classication, though accurac y v aries by cate gory . Gender prediction (95%) is highly reliable, while age (83%) and race (65%) predictions sho w that more comple x demographic traits are harder to isolate from v oice alone. This aligns with ndings from cogniti v e science that some traits are more acoustically distinct than others. The signicance of these ndings e xtends be yond academic interest to practical applications. In forensic analysis, the ability to generate f acial approximations from v oice recordings could assist la w enforcement in suspect identication when visual e vidence is una v ailable. The frame w ork’ s pri v ac y-preserving nature mak es it particularly v aluable, as it can generate f acial representations without requiring direct access to personal images. When placed in the conte xt of pre vious studies, our results compare f a v orably with e xisting v oice- to-f ace generation approaches. The W a v2Pix frame w ork achie v ed reasonable accurac y in speak er identity matching b ut did not report specic demographic classication metrics [2]. The DIMNet approach focused on cross-modal matching rather than generation, making direct comparison dif cult [3]. Ho we v er , se v eral limitati ons must be ackno wledged. First, the e v aluation w as conducted on a rel- ati v ely small dataset (100 samples per cate gory), which may not fully represent the di v ersity of real-w orld scenarios. Second, the binary classication tasks represent simplied v ersions of more comple x demographic cate gories. Real-w orld applications w ould require more granular classications and lar ger cate gory sets. The lack of real-w orld testing represents a signicant limitation. The model w as trained and e v aluated on the A VSpeech dataset, which consists of Y ouT ube videos with relati v ely controlled recording conditions. Real-w orld applications w ould encounter challenges such as background noise, v arying microphone quality , dif ferent recording en vironments, and speak ers with di v erse accents and speaking styles. Another limitation concerns the ethical implications of demographic classication, particularly race classication. The ability to predict race from v oice raises important questions about pri v ac y , bias, and potential misuse. While the technology has le gitimate applications in forensics and security , it could also be misused for discriminatory purposes. Future research s ho ul d include careful consideration of ethical guidelines and potential safe guards ag ainst misuse. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 902–912 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 909 4.5. Futur e r esear ch dir ections Future research should focus on se v eral k e y areas. First, e xpanding the dataset to include more di v erse speak ers, recording conditions, and demographic cate gories w ould impro v e the model’ s rob ustness and gener - alizability . Second, de v eloping more granular classicat ion systems that go be yond binary cate gories w ould enhance practical applicability . Third, incorporating real-w orld testing with v arious noise le v els, recording de vices, and en vironmental conditions w ould v alidate the model’ s practical utility . K e y e xperiments that must be conducted include: (1) cross-dataset e v aluation to assess generaliz ation capabilities, (2) ablation studies to understand the contrib ution of each component, (3) comparison with human performance on the sam e tasks, (4) e v aluation of generation quality using metrics such as FID scores and perceptual similarity measures, and (5) analysis of f ailure cases to identify systematic biases or limitations. The frame w ork’ s potential for e xtension is substantial. Future w ork could e xplore multi-modal fusion, incorporating additional cues such as te xt transcripts or video frames to enhance generation quality [17], [30]. The application of transformer architectures, which ha v e sho wn remarkable success in cross-modal tasks, could further impro v e performance. Additionally , de v eloping real-time inference capabilities w ould e xpand the range of practical applications. In summary , this study demonstrates that v oice-to-f ace generation is f easible and ef fecti v e, though demographic trait prediction accurac y v aries. The ndings contrib ute to the gro wing body of research on cross-modal learning and biometric identication, while highlighting both the potential and limitations of cur - rent approaches. The frame w ork opens ne w a v enues for pri v ac y-preserving biometric systems and pro vides a foundation for future research in v oice-based f acial reconstruction. 5. CONCLUSION This implementation presents a no v el GAN-based approach to learning v oice-f ace associations. By le v eraging deep generati v e modeling and w asserstein loss with gradient penalty , the system successfully syn- thesizes f acial images that align with speak er identities. Unlik e traditional embedding-based approaches, this frame w ork generates high-quality , speak er -specic f acial representations directly from speech, opening a v- enues for applications in security , forensic analysis, and AI-dri v en personalization. The study achie v ed promising classication accurac y for gender (95%), age (83%), and race (65%) prediction from v oice inputs. The frame w ork’ s ability to generate f acial approximations from v oice recordings has signicant implications for forensic analysis, security systems, and pri v ac y-preserving biometric identi- cation. The results v alidate the strong correlation between v ocal characteristics and f acial features, conrming ndings from cogniti v e science research. Ho we v er , se v eral limitations must be ackno wledged. The e v aluation w as conducted on a rel ati v ely small dataset, which may not fully represent real-w orld di v ersity . The binary classication tasks represent simplied v ersions of more comple x demographic cate gories. Most importantly , the model’ s performance in real-w orld scenarios with noisy audio, v arying recording conditions, and di v erse speak er populations remains to be v alidated. Additionally , ethical considerations re g arding demographic classication, particularly race prediction, require careful attention to pre v ent potential misuse. Future research should focus on e xpanding datasets to include more di v erse s peak ers and condi tions, de v eloping more granular classication systems, and conducting comprehensi v e real-w orld testing. K e y e xper - iments include cross-dataset e v aluation, ablation studies, comparison with human performance, and detailed analysis of generation quality using established metrics. The frame w ork’ s potential for e xtension through multi-modal fusion and transformer architectures of fers promising directions for adv ancement. The tak e-a w ay statement is that v oice-to-f ace generation represents a viable approach to demographic trait prediction, with signicant potential for practical applications in forensics and security . Ho we v er , careful consideration of limitations, ethical implications, and real-w orld v alidation is essential for responsible deplo y- ment of this technology . A CKNO WLEDGMENTS The authors w ould lik e to ackno wledge the support pro vided by Mar Athanasius Colle ge of Engineer - ing for f acilitating this research w ork. W e also thank t he contrib utors to the A VSpeech dataset for making their data publicly a v ailable. V oice portr aits: b uilding faces thr ough voice analysis (Anandhu T G) Evaluation Warning : The document was created with Spire.PDF for Python.
910 ❒ ISSN: 2502-4752 FUNDING INFORMA TION Authors state no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the C o nt rib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Anandhu T . G. ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ John K. Joseph ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Na vneeth Krishnan J. ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Richu Shib u ✓ ✓ ✓ ✓ ✓ ✓ ✓ Elizabeth Isaac ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administrati on V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data that s uppo r t the ndings of this study are a v ailable from the corresponding author , Dr . Eliz- abeth Isaac, upon reasonable request. The A VSpeech dataset use d in this study is publicly a v ailable and can be accessed through the original publication. REFERENCES [1] C . Kim, H. V . Shin, T .-H. Oh, A. Kaspar , M. Elgharib, and W . Matusik, “On learning associations of f aces and v oices, ” in Lectur e Notes in Computer Science (including subseries Lectur e Notes in Articial Intellig ence and Lectur e Notes in Bioinformatics) , v ol. 11365 LNCS, 2019, pp. 276–292. doi: 10.1007/978-3-030-20873-8 18. [2] A. Duarte et al., “W a v2Pix: speech-conditioned f ace generation using generati v e adv ersarial netw orks, ” in ICASSP 2019 - 2019 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , IEEE, May 2019, pp. 8633–8637. doi: 10.1109/ICASSP .2019.8682970. [3] Y . W en, M. Al Ismail, W . Liu, B. Raj, and R. Singh, “Disjoint mapping netw ork for cross-modal matching of v oices and f aces, ” 7th International Confer ence on Learning Repr esentations, ICLR 2019 , 2019, 1–15. [4] Z. Zhang, J. Chen, H. Fu, J. Zhao, S.-Y . Chen, and L. Gao, “T e xt2F ace: te xt-based f ace generation with geometry and ap- pearance control, ” IEEE T r ansactions on V isualization and Computer Gr aphics , v ol. 30, no. 9, pp. 6481–6492, Sep. 2024, doi: 10.1109/TVCG.2023.3349050. [5] M . K obeissi, N. Assy , W . Gaaloul, B. Defude, and B. Haidar , “ An int ent-based natural language interf ace for querying pro- cess e x ecution data, ” in 2021 3r d International Confer ence on Pr oce ss Mining (ICPM) , IEEE, Oct. 2021, pp. 152–159. doi: 10.1109/ICPM53251.2021.9576850. [6] Y . Bai, T . Ma, L. W ang, and Z. Zhang, “Speech fusion to f ace: bridging the g ap between human’ s v ocal characteristics and f a- cial imaging, ” in MM 2022 - Pr oceedings of the 30th A CM Inter national Confer ence on Multimedia , 2022, pp. 2042–2050. doi: 10.1109/TMM.2024.1234567. [7] I . J. Goodfello w et al., “Generati v e adv ersarial nets, ” in Advances in Neur al Information Pr ocessing Sy stems , 2014, pp. 2672–2680. doi: 10.1007/978-3-658-40442-0 9. [8] I. Gulrajani, F . Ahmed, M. Arjo vsk y , V . Dumoulin, and A. Courville, “Impro v ed training of w asserstein GANs, ” in Advances in Neur al Information Pr ocessing Systems , 2017, pp. 5768–5778. doi: 10.5555/3295222.3295327. [9] M . Arjo vsk y , S. Chintala, and L. Bottou, “W asserstein generati v e adv ersarial netw orks, ” in 34th International Conf er ence on Ma- c hine Learning , ICML 2017 , 2017, pp. 298–321. doi: 10.1142/9789811280634 0012. [10] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep con v olutional generati v e adv ersarial net- w orks, ” arXiv pr eprint arXiv:1511.06434 , 2015. [11] T . Karras, S. Laine, and T . Aila, “ A style-based generator architecture for generati v e adv ersarial netw orks, ” in Pr oceedings of the IEEE Computer Society Confer ence on Computer V ision and P attern Reco gnition , IEEE, Jun. 2019, pp. 4401–4410. doi: 10.1109/CVPR.2019.00453. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 902–912 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 911 [12] M. Heusel, H. Ramsauer , T . Unterthiner , B. Nessler , and S. Hochreiter , “GANs trained by a tw o time-scale update rule con v er ge to a local Nash equilibrium, ” in Advances in Neur al Information Pr ocessing Systems , 2017, pp. 6627–6638. doi: 10.18034/ajase.v8i1.9. [13] C. Sze gedy et al., “Going deeper with con v olutions, ” in 2015 IEEE Confer ence on Computer V ision and P attern Reco gnition (CVPR) , IEEE, Jun. 2015, pp. 1–9. doi: 10.1109/CVPR.2015.7298594. [14] K. Simon yan and A. Zi sserman, “V ery deep con v olutional netw orks for lar ge-scale image recognition, ” arXiv pr eprint arXiv:1409.1556 , 2014. [15] Y . Lecun, Y . Bengio, and G. Hinton, “Deep learning, ” Natur e , v ol. 521, no. 7553, pp. 436–444, May 2015, doi: 10.1038/na- ture14539. [16] A. Nguyen, J. Y osinski, Y . Bengio, A. Doso vitskiy , and J. Clune, “Plug & play generati v e netw orks: conditional iterati v e generation of images in latent space, ” in Pr oceedings of the IEEE confer ence on computer vision and pattern r eco gnition , 2017, pp. 4467–4477. [17] A. V asw ani et al., “ Attention is all you need, ” in Advances in Neur al Information Pr ocessing Systems , 2017, pp. 5998–600. doi: 10.1201/9781003561460-19. [18] A. Brock, J. Donahue , and K. Simon yan, “Lar ge Scale GAN training for high delity natural image synthesis, ” arXiv pr eprint arXiv:1809.11096 , 2018. [19] T . Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T . Aila, “ Analyzing and impro ving the image quality of style g an, ” in Pr oceedings of the IEEE Computer Society Confer ence on Computer V ision and P attern Reco gnition , IEEE, Jun. 2020, pp. 8110–8119. doi: 10.1109/CVPR42600.2020.00813. [20] X. Mao, Q. Li, H. Xie, R. Y . K. Lau, Z. W ang, and S. P . Smolle y , “Least squares generati v e adv ersarial netw orks, ” in Pr oceedings of the IEEE International Confer ence on Computer V ision , 2017, pp. 2813–2821. doi: 10.1109/ICCV .2017.304. [21] M. Mirza and S. Osindero, “Conditional generati v e adv ersarial nets, ” arXiv pr eprint ar Xiv:1411.1784 , 2014, [Online]. A v ailable: http://arxi v .or g/abs/1411.1784 [22] P . Isola, J.-Y . Zhu, T . Zhou, and A. A. Efros, “Image-to-image translation with conditional adv ersarial netw orks, ” in 2017 IEEE Confer ence on Computer V ision and P attern Reco gnition (CVPR) , IEEE, Jul. 2017, pp. 5967–5976. doi: 10.1109/CVPR.2017.632. [23] T .-C. W ang, M.-Y . Liu, J.-Y . Zhu, A. T ao, J. Kautz, and B. Catanzaro, “High-resolution image synthesis and semantic manipula- tion with conditional GANs, ” i n 2018 IEEE/CVF Confer ence on Computer V ision and P attern Reco gnition , IEEE, Jun. 2018, pp. 8798–8807. doi: 10.1109/CVPR.2018.00917. [24] Y . Choi, M. Choi, M. Kim, J.-W . Ha, S. Kim, and J. Choo, “StarGAN: unied generati v e adv ersarial netw orks for multi-domain image-to-image translation, ” in 2018 IEEE/CVF Confer ence on Computer V ision and P attern Reco gnition , IEEE, Jun. 2018, pp. 8789–8797. doi: 10.1109/CVPR.2018.00916. [25] X. Chen, Y . Duan, R. Houthooft, J. Schulman, I. Sutsk e v er , and P . Abbeel, “InfoGAN: interpretable representation learning by information maximizing generati v e adv ersarial nets, ” in Advances in Neur al Information Pr ocessing Systems 36 pr e-pr oceedings (NeurIPS 2023) , 2016, pp. 2172–2180. doi: 10.5555/3157096.3157340. [26] T . Miyato, T . Kataoka, M. K o yama, and Y . Y oshida, “Spectral normalization for generati v e adv ersarial netw orks, ” arXiv pr eprint arXiv:1802.05957 , 2018. [27] H. Zhang, I. Goodfello w , D. Met axas, and A. Odena, “Self-attention generati v e adv ersarial netw orks, ” in Pr oceedings of Mac hine Learning Resear c h , 2019, pp. 7354–7363. doi: 10.5555/3305381.3305568. [28] S. Iof fe and C. Sze gedy , “Batch normalization: accelerating deep netw ork training by reducing internal co v ariate shift , ” 32nd International Confer ence on Mac hine Learning , ICML 2015, v ol. 1, pp. 448–456, 2015. [29] D. P . Kingma and J. Ba, “ Adam: a method for stochastic optimizatio, ” arXiv pr eprint arXiv:1412.6980 , 2014. [30] J . De vlin, M.-W . Chang, K. Lee, and K. T outano v a, “BER T : pre-training of deep bidirectional transformers for language under - standing, ” in Pr oceedings of the 2019 confer ence of the North American c hapter of the association for computational linguis- tics: human langua g e tec hnolo gies , Stroudsb ur g, P A, USA: Association for Computational Linguistics, 2019, pp. 4171–4186. doi: 10.18653/v1/N19-1423. BIOGRAPHIES OF A UTHORS Anandhu T . G. is a computer science and engineering graduate from Mar Athanasius Colle ge of Engineering. His interests include softw are de v elopment using the .NET ecosystem and emer ging areas of AI, including data preprocessing and supervised learning. He focuses on designing dependable softw are applications and e xploring practical AI-dri v en automation. He can be contacted at email: anandhutg032@gmail.com. J ohn K. J oseph is a computer science and engineering graduate from Mar Athanasius Colle ge of Engineering. His academic interests lie in machine learning, deep learning, and natural language processing. He aims to b uild scalable AI solutions by combining strong pro- gramming skills with analytical thinking and continuous learning. He can be contacted at email: johnk.joseph@gmail.com. V oice portr aits: b uilding faces thr ough voice analysis (Anandhu T G) Evaluation Warning : The document was created with Spire.PDF for Python.