DeCiphr icon

An AI-assisted tool for
linguistic landscape analysis

Systematic, auditable coding of signs and multimodal data. Collect in the field by phone or import an existing corpus at your desk — you set the framework, review every result, and export a full audit trail.

Download on theApp Store
Coming soon onGoogle Play
Use onDesktop App
No account required  ·  All data stays on your device  ·  Built for academic research

A field work companion and data processing application
for linguistic landscape research

Automatic Data Processing

Upload any number of images and DeCiphr does the groundwork: every item is assigned a unique admin ID, sorted into a clear order, and any image containing multiple distinct visual items is split so each is independently numbered and catalogued. Human faces are detected and blurred on your device by default before any image is transmitted or stored — an ethical safeguard for research in public space, which you can turn off when appropriate (for example, public figures on public signage). Your corpus is organized and ready before any analysis begins.

Custom Frameworks

Define your own coding scheme. Works with any established or original framework. Share as a portable .sigframe file with collaborators.

AI-Assisted Coding

Analysis runs on Anthropic's Claude vision API: the model reads each image and proposes codes from the framework you defined. You review and edit every output, and the exact model, prompt, and raw response are logged for each sign. It runs in strict coding mode, so the AI applies only the codes you defined.

GPS Field Mapping

Capture public signs in the field with automatic GPS tagging. View all your signs plotted on an interactive map, organized by project.

Analytic Memos & Audit Trail

Write interpretive field notes directly on each sign. Every AI prompt and raw response is logged alongside your edits — giving you a complete, exportable record of every analytical decision for full methodological transparency.

Export for NVivo & Atlas.ti

Export structured CSV files that import into NVivo, Atlas.ti, and MAXQDA, ready for your existing research workflow.

Research Methodological Considerations

Reproducibility

Full AI Audit Trail

Every analysis leaves a complete paper trail: the exact prompt sent to the AI, the full raw response, and a timestamped record of every edit you made to the output. All of it is exportable as part of your dataset.

Exact system prompt logged per sign — so your analysis is fully replicable
Raw AI output saved alongside your final coded values
Researcher edits tracked separately from AI-generated data
First 500 characters of each prompt included in CSV export

From photographs
to a coded corpus

1

Create or import a framework

Build your coding scheme from scratch, or load an established framework shared as a .sigframe file. Define codes, values, and a master prompt for the AI.

2

Import your data

Capture signs in the field with GPS tagging, or import screenshots and digital images. DeCiphr automatically assigns each item a unique admin ID, sorts your corpus into order, and separates any image containing multiple visual items into independently numbered entries. Any human faces are detected and blurred on-device by default before storage or transmission.

3

Run AI coding

DeCiphr sends each image to Claude Vision and transcribes and translates language visible on the image. Then it proposes codes from your framework. You then review, edit, and annotate every result — and the prompt and raw response are saved to the audit trail.

4

Export structured data

Export a complete CSV with all coded values, languages detected, GPS coordinates, analytic memos, and metadata.

Designed for linguistic landscape
and visual data research

DeCiphr is built by an active field researcher, and its design is grounded in the methodological demands of linguistic landscape research — from systematic data collection in public spaces to structured, framework-driven analysis. Whether working in the field with a phone or processing a corpus at a desk, the workflow is the same.

How the AI works

Understand exactly what the AI is doing — and what it isn't. Here are the answers to some of the questions that matter methodologically.

DeCiphr runs in strict coding mode — a core methodological safeguard designed for rigorous research.

Strict coding mode The AI is instructed to assign only the values defined in your framework codes, and not to introduce observations, categories, or terminology beyond what you have defined. If the image does not clearly match a code, it is instructed to output "uncodable" rather than guess. You review and edit every result, so the final coding reflects your framework — not the AI's assumptions.
It is instructed, not trained. Training would mean permanently modifying the underlying AI model — which is not possible here. Instead, DeCiphr uses prompt engineering: your framework, codes, and definitions are passed to the AI as precise instructions at the moment of analysis. The AI reads your image using its full visual capabilities, but what it is permitted to report is controlled by your framework and the mode you have selected. This distinction is important for methodological transparency in research publications.
Yes — this is one of DeCiphr's core use cases. The built-in frameworks are grounded in established analytical approaches in linguistic landscape studies, multimodal discourse analysis, and related fields. For any other analytical framework, you can create a custom scheme: enter the original codes and their definitions, write a master prompt that situates the analysis within the appropriate analytical context, and enable strict coding mode to ensure the AI applies only those codes. The .sigframe file format also allows you to share your configured framework with other researchers so the analytical procedure is fully replicable.
All your data — photos, coded results, frameworks, and analytic memos — is stored exclusively on your device. Nothing is uploaded to any server operated by the app developer. When you run an AI analysis, the image is sent to Anthropic's API (Claude) solely to perform that analysis. Anthropic does not use API inputs or outputs to train its models, and automatically deletes them within 30 days, with limited exceptions for legal compliance and abuse monitoring. The app developer has no access to your images, data, or frameworks at any time.

Stay Updated

Get notified when new frameworks, features, and research using DeCiphr are published. No spam — just meaningful updates for researchers.

✓ You're subscribed! We'll notify you when new frameworks and features are released.

No spam. Unsubscribe any time. Your email is never shared or sold.

Linguistic landscape (LL) research has always been shaped by the tools available for capturing, organizing, and interpreting public signage, and each technological advance has widened what researchers are able to document and ask. This is evident in highly multilingual settings, where the field researcher may not be competent in all the languages available in the public space, leading to either hiring translators or adopting digital tools able to recognize unfamiliar scripts.

One of the first technological contributions addressed location and collection. Barni and Bagna’s (2009) MapGeoLing used smartphone GPS to locate signs on a map, enabling quantitative analysis of linguistic distribution across space. Mobile applications then standardized the photographic capture and geotagging of signs: LinguaSnapp (Gaiser & Matras, 2021) built a structured LL archive, while Lingscape (Purschke, 2017) adopted a citizen-science model in which members of the public crowdsource sign data.

Another contribution concerns analysis. Mignone (2025) used the qualitative data analysis package ATLAS.ti to build a coding system grounded in geosemiotics, uploading photographs and assigning codes manually alongside interview transcripts and fieldnotes within a single project. This integration supports queries and co-occurrence analysis across a dataset, but the coding itself remains a manual, researcher-driven act. More automated approaches have come from geography and urban studies: Hong (2020) applied computer vision and machine learning to extract and code text from online street-level imagery in order to demonstrate linguistic diversity across urban space.

Most recently, attention has turned to artificial intelligence. Voss (2024) proposes generative AI for indexing, hypothesis generation, and interpretive support, arguing that it should function as a research assistant in dialogue with the researcher rather than as a replacement for analytical judgment: “Researchers must use AI technologies responsibly and consider them as tools to augment human expertise rather than replace it” (p. 419). This framing sharpens the central methodological question for the field: what can AI do with LL data? At a minimum, it can identify signs, transcribe, and translate text, and it can process data according to researcher-defined coding—provided its interpretation is constrained to predetermined codes rather than left open.

DeCiphr is the tool I have developed in response to this question. It is a smartphone and web application for coding LL data in which the researcher creates or imports a coding framework and the AI codes within it. As a mobile application, it is designed as a fieldwork companion in multilingual contexts, processing signs as they are collected—translating and coding on site, in precisely the conditions my own project demands. Each framework defines its codes, definitions, and permissible values, and is organized into two layers.

Taking geosemiotics (Scollon & Wong Scollon, 2003) as an example, Layer 1 holds the deductive codes the AI assigns: seven geosemiotic codes including code preference, inscription, emplacement, represented participants, modality, composition, and interactive participants. Here the AI records only what is visually present—languages, layout, font, and composition—and returns “Uncodable” wherever the visual evidence does not warrant a confident assignment. Layer 2 holds interpretive questions that the researcher alone answers, drawing on fieldwork presence, community knowledge, and ethnographic judgment. The AI cannot answer these by design, and the researcher retains authority throughout, able to edit both the processing and the Layer 1 output.

This division clarifies both the potential and the limitations of AI for qualitative LL research. Its potential is considerable: it can manage large data, support translation across unfamiliar languages, and give the researcher more time to concentrate on interpretation, while its guiding questions can scaffold students and less experienced researchers. However, it can be argued that the same guiding questions that scaffold early researchers may also produce shallow or restrictive responses. Another issue to consider is agency: allowing the researcher to edit AI output distributes agency and accountability across human and machine in ways that are difficult to disentangle. It is also important to remember that large language models are not trained evenly across cultures, and this unevenness affects coding, transcription, and translation alike.

Taken together, these tools trace a clear trajectory from locating and collecting signs, to organizing and querying them, to coding them. DeCiphr extends that trajectory into AI-assisted coding while deliberately preserving the researcher’s interpretive authority. Treated as an enhancement to the researcher’s experience, AI can strengthen the reflexive, qualitative work at the heart of LL research without displacing the researcher’s role.

References

Barni, M., & Bagna, C. (2009). A mapping technique and the linguistic landscape. In E. Shohamy & D. Gorter (Eds.), Linguistic landscape: Expanding the scenery (pp. 126–140). Routledge.

Gaiser, L., & Matras, Y. (2021). LinguaSnapp Manchester: Multilingualism revealed. University of Manchester.

Hong, S.-Y. (2020). Linguistic landscapes on street-level images. ISPRS International Journal of Geo-Information.

Mignone, O. (2025). Vital signs: Tibetan in the linguistic landscape of Jackson Heights [Doctoral dissertation, City University of New York].

Purschke, C. (2017). Crowdsourcing the linguistic landscape: Introducing Lingscape. Linguistik Online, 85(6).

Scollon, R., & Wong Scollon, S. (2003). Discourses in place: Language in the material world. Routledge.

Voss, E. (2024). Artificial intelligence and linguistic landscape research: Affordances, challenges & considerations. Linguistic Landscape, 10(4), 400–424.

Salem M. Al-Ajmi
PhD Researcher · University of Reading
University of Reading, UK
s.m.f.alajmi@pgr.reading.ac.uk

How to cite: Al-Ajmi, S. M. (2026). DeCiphr: An AI Research Assistant [Computer software]. https://deciphr.tech/

Begin your corpus

No account required. All data is stored locally on your device.