cs.CLOct 12, 2026

The interpreter as developer: An autoethnographic design study of building a computer-assisted interpreting tool through human-LLM co-creation

Authors: Zeming Zhao

Organizations: College of Foreign Languages, Capital Normal University, Beijing, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Real-time AI captioning now serves audiences at multilingual events, yet computer-assisted interpreting (CAI) tools are still built for interpreters, rarely by them. This article reports an autoethnographic design study in which a professional interpreter with no programming training built a local-first CAI system through expert-led 'vibe coding', co-creating with an LLM architect and a coding agent. Analysis of the complete development dialogues shows how tacit professional knowledge, from segmentation to booth infrastructure, was externalized into design after generic defaults failed in use. In a demanding public scientific event, three mechanisms of in-booth appropriation emerge under a firmly human-led division of labor; telemetry from a first fully instrumented assignment shows a fully local chain sustaining a median translation latency of 0.94 s across 971 utterances. The case is read as lead-user innovation newly enabled by generative AI, with implications for the technological literacy of interpreters and the future of interpreter training.

Figures & tables

Figure 1: The triadic co-creation workflow. The domain expert originates requirements and holds veto power; the architect LLM converts practitioner complaint into structured build instructions; the coding agent implements. Thin and dashed arrows mark the two bypass paths (direct small-fault reports; context re-establishment between platforms). Commercial substrates changed mid-project (Gemini 3.1 Pro →\rightarrow GLM-5.3; Cursor →\rightarrow ZCode); the roles did not.
#Tacit knowledge (practice)Design requirementMechanismEvidence
1Console 3.5 mm analogue output is the only reliable booth interface; monitoring untouchableBridge console audio without degrading monitoringUSB audio bridge; analogue split, isolated monitoring branchDialogues
2Consumer pipelines ‘improve’ the signal in ways that corrupt a console feedClean pass-through; level control without a mixerBrowser DSP (AEC/NS/AGC) disabled at capture; non-linear gainCode audit
3Booth connectivity is weak; events are confidentiality-boundNo in-booth cloud dependenceLocal STT trio + local quantized LLM; optional cloud bypass; auto-degradeCode audit
4Support arriving after my output is uselessLatency budget within/near EVSHot-switchable STT per speaker; whole-utterance finalization; short prompt historyCode; telemetry (Figure 2)
5Fixed slicing severs syntax; over-eager silence detection shatters wordsSegmentation semantic-first, time-capped as fallbackPunctuation trigger + tunable pause (0.3–2.0 s) + hard cap (5–30 s) + minimum-speech guardEpisode 1
6Small models ‘parrot’ under mixed-language noise; blunt script-filtering destroys acronymsOne-way direction, source-language terms unharmedCharacter-ratio probe + negative instructions + echo filter ( ≥\geq 85% overlap discarded)Episode 2
Table 1: Mapping tacit interpreting knowledge to design decisions.
Figure 2: Translation latency on the instrumented assignment of 28 August 2026 (n = 971 utterances; SenseVoice recognition, Qwen2.5-7B translation, fully local chain). Histogram (0.25 s bins) with cumulative distribution; one 46 s cold start excluded. The median (0.94 s) falls within the interpreter’s plausible ear–voice span window; the p90–p95 tail (2.16–2.43 s) marks where support risks arriving after the interpreter has committed to a rendering.
More in this topic

Human-AI Co-Creation

Human-AI co-creation studies how people and AI jointly develop ideas, artifacts, and decisions while preserving meaningful human agency.

See all 154 papers

Explore similar work

CardsList

  1. DeepTrans Studio: Turning Expert Interventions into Shared Team Knowledge in Agentic Translation Workflows

    Jun 29, 2026Ziyang Lian, Qingya Zhang, Hao Wang +5Agentic WorkflowsHuman-AI Collaboration

    School of Computer Engineering and Science, Shanghai University, China · School of Foreign Languages, East China Normal University, China · School of Foreign Studies, Shanghai University, China +1

  2. Key Considerations for Domain Expert Involvement in LLM Design and Evaluation: An Ethnographic Study

    Feb 16, 2026Annalisa Szymanski, Oghenemaro Anuyah, Toby Jia-Jun Li +1LLM EvaluationHuman-in-the-Loop Evaluation

    Computer Science and Engineering, University of Notre Dame, Notre Dame, IN, USA · Microsoft Corporation, Redmond, WA, USA

  3. Bridging the Usability Gap: Lessons from Interpreting Studies for Machine Interpreting Design

    Jun 14, 2026Claudio FantinuoliSpeech TranslationSpeech Translation Evaluation

    University of Mainz