# Talking Face Generation in Socio-Technical Systems: A Systematic Review of Deep Learning Architectures, Deployment Contexts, and Governance Frameworks

**Authors:** Ridwan Salahudeen (Department of Computer Science, Ahmadu Bello University, Zaria, Nigeria) — ORCID 0009-0000-1097-0906; Mathias Fonkam (College of Information Sciences and Technology, Pennsylvania State University, University Park, PA, USA) — ORCID 0000-0002-2776-1462; Narasimha Rao Vajjhala (Department of Computer Science, American University in Bulgaria, Blagoevgrad, Bulgaria) — ORCID 0000-0002-8260-2392; Sahalu Balarabe Junaidu (Department of Computer Science, Ahmadu Bello University, Zaria, Nigeria); Aliyu Garba (Department of Computer Science, Ahmadu Bello University, Zaria, Nigeria) — ORCID 0000-0003-1875-6167
**Type:** Systematic review (open access)
**Source:** Systems, 14(10), Article 1198
**Published:** 2026-09-24
**DOI:** https://doi.org/10.3390/systems14101198
**Canonical page:** https://www.narasimharao.net/research/talking-face-generation-socio-technical-systems-systematic-review/
**Indexing:** Scopus Q1 (CiteScore 5.4) · Web of Science SSCI Q1 · Impact Factor 3.8 (JCR 2025)
**Keywords:** talking face generation; socio-technical systems; deepfake governance; systematic literature review; responsible AI; human-AI interaction
**Methodology:** Systematic literature review (Kitchenham and Charters; reported to PRISMA 2020) of 92 primary studies published 2016–2025, screened from 18,343 records in six repositories over four search iterations, with a five-criterion quality assessment and qualitative narrative synthesis across six research questions.

## Research summary

- **The Problem:** Talking face generation can now synthesise photorealistic, real-time speaking video from a single image and an audio clip and is entering healthcare, education, identity management and public communication, yet no prior review was at once systematic, current to 2024–2025, and comprehensive across its technical, ethical and socio-technical dimensions.
- **The Methodology:** Systematic literature review (Kitchenham and Charters; reported to PRISMA 2020) of 92 primary studies published 2016–2025, screened from 18,343 records in six repositories over four search iterations, with a five-criterion quality assessment and qualitative narrative synthesis across six research questions.
- **The Core Finding:** Across 92 primary studies (2016–2025), Diffusion Transformers and Gaussian Splatting account for about 36% of studies dated 2024 or later while GAN use has fallen below 3%; training data are demographically skewed and under-documented, and the dominant metrics (PSNR, SSIM, SyncNet) do not measure what deployment requires. Only one of the 92 systems embeds any technical safeguard — a non-compliant post hoc watermark — and none implements consent or provenance, even though EU AI Act Article 50 transparency obligations have applied since 2 August 2026.
- **The Citation:** Salahudeen, R., Fonkam, M., Vajjhala, N. R., Junaidu, S. B., & Garba, A. (2026). Talking Face Generation in Socio-Technical Systems: A Systematic Review of Deep Learning Architectures, Deployment Contexts, and Governance Frameworks. Systems, 14(10), 1198. https://doi.org/10.3390/systems14101198

## Summary in detail

**Question.** How have deep learning architectures for talking face generation evolved since 2016, how well do their training data, evaluation metrics and safeguards match what real deployments require, and what governance would make their use in healthcare, education, identity management and public communication responsible?

**Finding.** Synthesising 92 primary studies published between 2016 and 2025, the review finds that talking face generation has passed through five architectural generations, with Diffusion Transformers and Gaussian Splatting now accounting for about 36% of studies dated 2024 or later and GAN use below 3%. Training data are demographically skewed and under-documented, the dominant metrics (PSNR, SSIM, SyncNet) do not measure what deployment requires, and only one of the 92 systems embeds any technical safeguard — a post hoc watermark that is not C2PA-compliant — while none implements consent or provenance.

**Why it matters.** Talking face generation now produces photorealistic, real-time speaking video from a single image and an audio clip and is entering healthcare, education, identity management and public communication, yet the reviewed state of the art satisfies none of the EU AI Act Article 50 transparency obligations that have applied since 2 August 2026. The authors argue that the gap is structural rather than a failure of individual researchers — publication incentives reward benchmark optimisation alone — so venues, funders and regulators need to make provenance, consent and demographic auditing conditions of publication and deployment. The RTFG-STS Framework and three proposed deployment-fitness metrics give researchers and deployers a starting point that still requires empirical validation.

## Key findings

1. The review synthesised 92 primary studies published between January 2016 and December 2025, selected from 18,343 records retrieved across six repositories in four search iterations (March 2023 – March 2026) under Kitchenham’s guidelines and PRISMA 2020.
2. Talking face generation has passed through five architectural phases: LSTM and RNN lip-sync models (2016–2018), GANs (2019–2022), transformers and Neural Radiance Fields (2021–2023), diffusion models (2023–2024), and Diffusion Transformers (DiT) with Gaussian Splatting (2024–2025).
3. Of the 41 studies dated 2024 or later, 15 (36.6%) are DiT or Gaussian Splatting systems and only one (2.4%) is GAN-based; together, DiT and Gaussian Splatting have resolved the field’s historic trade-off between latency and quality.
4. Training data are demographically skewed and under-documented: VoxCeleb2 underpins 60 studies (65.2%), 14 of the 19 named corpora are English-only, only CREMA-D (used in a single study) reports ethnicity statistics, and no included study reports a demographic audit of its data.
5. Evaluation still relies on FID (65.2% of studies), SyncNet-derived synchrony scores (55.4%), SSIM (46.7%) and PSNR (37.0%); PSNR, SSIM and SyncNet measure none of identity preservation, cross-cultural naturalness, trustworthiness or detectability, and the identity metric CSIM appears in only 15.2% of studies.
6. Only 8 of the 92 studies (8.7%) contain any ethical content, and exactly one (1.1%) integrates a technical safeguard — a post hoc frequency-domain watermark that is not C2PA-compliant and untested against compression. No study implements a consent protocol or cites the EU AI Act, the TAKE IT DOWN Act, C2PA or any other governance instrument as a design constraint.
7. More than half of the studies (48, 52.2%) specify no deployment domain; among those that do, entertainment leads (18 studies), followed by education (9), healthcare (7), identity management (5), governance (3) and accessibility (2).
8. The authors propose the RTFG-STS Framework, which adds a sixth principle — systemic embeddedness — to the five AI4People principles and operationalises them through five mechanisms (provenance and marking, risk-tiered consent, socio-technical impact assessment, a deepfake detectability audit, and human oversight), together with three deployment-fitness metrics: the Perceived Trustworthiness Score, the Cross-Cultural Authenticity Index and the Deepfake Detectability Rate — all offered as research constructs that require empirical validation.

## Study at a glance

| Item | Detail |
|---|---|
| Research questions | Six: architecture evolution; dataset bias and generalisation; evaluation metrics and deployment fitness; DiT and Gaussian Splatting versus GANs and NeRFs; ethical safeguards and regulatory alignment; TFG in socio-technical systems |
| Design | Systematic literature review following Kitchenham and Charters, reported to PRISMA 2020; qualitative narrative synthesis (no meta-analysis, because studies report incompatible metrics on incompatible test sets) |
| Sources | IEEE Xplore, ScienceDirect, ACM Digital Library, SpringerLink, Google Scholar with Web of Science, and arXiv; four search iterations between March 2023 and March 2026 |
| Screening | 18,343 records retrieved → 8,741 after deduplication → 448 abstracts screened by two reviewers (Cohen’s κ = 0.84) → 322 full texts quality-scored → 92 included |
| Corpus | 92 primary studies (January 2016 – December 2025): 56 conference papers, 24 journal articles and 12 preprints, spanning nine architecture families, 19 named datasets and 17 evaluation metrics |
| Main result | DiT and Gaussian Splatting ≈ 36% of studies from 2024 onward; 8 of 92 studies (8.7%) contain any ethical content; 1 embeds a safeguard; none implements consent or provenance |
| Contribution | The RTFG-STS Framework (six principles, five mechanisms) and three deployment-fitness metrics — PTS, CCAI and DDR — proposed as research constructs requiring empirical validation |

## Abstract

Talking face generation (TFG)—the synthesis of photorealistic speaking video from a portrait and an audio signal—has passed through five architectural generations since 2016, from Long Short-Term Memory (LSTM) lip-sync models through Generative Adversarial Networks (GANs), Neural Radiance Fields (NeRFs) and Diffusion Models to Diffusion Transformers (DiT) and Gaussian Splatting, and now enters healthcare, education, identity management and public communication in real time. This systematic review synthesises 92 primary studies (January 2016–December 2025), selected from 18,343 records across six repositories under Kitchenham’s guidelines and PRISMA 2020, addressing architecture evolution, dataset bias, evaluation metrics, paradigm shifts, ethical safeguards and socio-technical deployment. DiT and Gaussian Splatting account for roughly 36% of studies dated 2024 or later, and GAN use has fallen below 3%; training corpora are demographically skewed and under-documented, and the dominant metrics (PSNR, SSIM, SyncNet) do not measure what deployment requires. Only one of the 92 systems embeds any technical safeguard, a non-compliant post hoc watermark, and none implements consent or provenance, although the EU AI Act, the TAKE IT DOWN Act, China’s labelling Measures and C2PA v2.0 were all adopted within the review period. We propose three deployment-fitness metrics and the RTFG-STS Framework as research constructs requiring empirical validation.

## When this research may be relevant

This review may be relevant to researchers surveying talking face or talking head generation, audio-driven portrait animation and deepfake synthesis, and to work on dataset bias, demographic representation and evaluation metrics for generative video models. It also speaks to AI governance and regulation of synthetic media — content provenance (C2PA), watermarking, consent and EU AI Act Article 50 transparency obligations — and offers a socio-technical systems view of deploying generative avatars in healthcare, education, identity verification and public communication.

## Key terms

- **Talking face generation (TFG):** The synthesis of photorealistic, temporally coherent video in which a face speaks in synchrony with a given audio signal, typically from a single portrait image and an audio clip.
- **Diffusion Transformer (DiT):** A diffusion model in which the usual U-Net denoising network is replaced by a transformer operating over latent tokens, so that output quality scales with model size and compute.
- **Gaussian Splatting:** A 3D rendering approach that represents a head or scene as learned 3D Gaussians and draws them by fast differentiable rasterisation instead of volume rendering, which makes real-time talking-head synthesis possible.
- **Content provenance (C2PA):** A cryptographically signed manifest, specified by the Coalition for Content Provenance and Authenticity (C2PA v2.0), recording how a media asset was produced and bound to that asset, so that the authenticity and integrity of the claim can be verified.
- **Deepfake Detectability Rate (DDR):** A metric proposed in this review: the proportion of generated frames correctly classified as synthetic by an ensemble of state-of-the-art deepfake detectors, intended to operationalise the machine-detectability required by EU AI Act Article 50(2).
- **Systemic embeddedness:** The sixth principle the RTFG-STS Framework adds to the five AI4People principles: the design and governance of a generative system should reference the concrete socio-technical context in which it is deployed, not the model in isolation.

## Limitations

- Restricting the review to English-language publications underrepresents Chinese-language venues, although Chinese institutions are among the field’s most active contributors.
- Studies report incompatible metric subsets on incompatible test sets, so no meta-analysis was possible and differences in quality between architectures cannot be quantified from the published literature.
- The findings on ethical safeguards and socio-technical deployment (RQ5, RQ6) rest on a thinner evidence base that also draws on regulatory instruments, industry reporting and adjacent literature gathered purposively rather than under the review protocol.
- The three deployment-fitness metrics and the RTFG-STS Framework are research constructs that have not been validated against human raters or a detection benchmark, and are not compliance criteria.
- The regulatory analysis reflects the position as at September 2026 and should be re-checked against later EU AI Office guidance.

## How to cite

Salahudeen, R., Fonkam, M., Vajjhala, N. R., Junaidu, S. B., & Garba, A. (2026). Talking Face Generation in Socio-Technical Systems: A Systematic Review of Deep Learning Architectures, Deployment Contexts, and Governance Frameworks. Systems, 14(10), 1198. https://doi.org/10.3390/systems14101198

```bibtex
@article{salahudeen2026talking,
  title = {Talking Face Generation in Socio-Technical Systems: A Systematic Review of Deep Learning Architectures, Deployment Contexts, and Governance Frameworks},
  author = {Salahudeen, Ridwan and Fonkam, Mathias and Vajjhala, Narasimha Rao and Junaidu, Sahalu Balarabe and Garba, Aliyu},
  journal = {Systems},
  volume = {14},
  number = {10},
  pages = {1198},
  year = {2026},
  publisher = {MDPI},
  doi = {10.3390/systems14101198},
  url = {https://doi.org/10.3390/systems14101198}
}
```