Cortex LLM is the language model at the core of Argus. Its architecture is inspired by how the human brain perceives: it takes in many signals at once, chooses what to attend to, keeps context in working memory, predicts from prior knowledge and checks its own confidence, instead of translating one isolated sentence at a time.
THE PROBLEM
Today's translation systems are designed for individual users and process one sentence at a time with no running context. That works for a two-person chat, but fails in organizational meetings with many speakers, many languages and context that builds up over the whole meeting.
The system must know which language each person speaks without setup, and keep that identification stable instead of flipping back and forth.
Speakers in Asian organizations often mix English into Thai, Chinese or Japanese sentences.
Sentence-by-sentence translation loses pronouns, references to earlier points and the organization's own terminology.
Fast translation tends to be inaccurate; accurate translation tends to be too slow for live conversation.
Users can't see how the AI decides, so they hesitate to rely on it in important meetings.
Thai speech data from business meetings is still scarce compared with English and Chinese.
The human brain solves these problems every day.
A professional interpreter listens to many people, remembers context, guesses words they didn't hear clearly, and knows when they're unsure. Cortex asks whether those brain principles can be turned into an AI architecture.
THEORETICAL FRAMEWORK
Cortex doesn't simulate neurons. It takes working principles accepted in cognitive neuroscience and uses them as an engineering metaphor to define each module's job, not as a claim to work like the brain biologically.
| BRAIN PRINCIPLE | IN THE BRAIN | IN CORTEX |
|---|---|---|
| Multisensory integration | Merges signals from several senses into one percept | Combines words, tone, rhythm and emotion into one representation of speech |
| Voice recognition | Dedicated areas recognize individual voices | Speaker embeddings (ECAPA-TDNN) bound to each speaker's identity and language |
| Selective attention Cocktail party effect | Tunes in to a target voice amid noise | Attention that weights the main speaker and filters nearby voices |
| Working memory Baddeley | Holds information temporarily for ongoing processing | Meeting Context Buffer keeps speakers, topics and references for the whole meeting |
| Predictive processing Friston | Predicts incoming input, then corrects from the error | Topic-aware Prior predicts terms and meaning from the topic and fixes unclear words |
| Language centers Wernicke / Broca | Comprehension and production are separate processes that work together | Separates the Comprehension layer from target-language Generation |
| Global workspace Baars | Important information is broadcast to many systems at once | A shared workspace that every module reads and writes |
| Long-term memory consolidation | Moves what matters from short-term to lasting memory | After the meeting, knowledge is summarized into the organization's knowledge base, with consent |
| Prefrontal executive function | Plans, weighs options and controls behavior | The Consult and Act modules |
| Metacognition | Knows whether it knows | Confidence estimation flags uncertain points for users to check |
ARCHITECTURE
Processing runs in four layers, from perception to action. Signals flow downward, while the Global Workspace sends context back to every layer: meeting memory helps translation, and the Predictive Prior helps perception fix words it didn't hear clearly.
Speech to text, voiceprint, tone and emotion
Main speaker, noise filtering, stable language binding
Meaning and intent, rendered in many languages at once
Advise, create tasks, summarize into Origami
Every layer reads and writes shared context here, in both directions.
Working Memory
Meeting context buffer: speakers, topics and references
Long-term Memory
Organizational knowledge and terminology, with consent
Predictive Prior
Anticipates terms and meaning from the current topic
Metacognition
Estimates confidence and shows uncertain points to the user
⇄ Connected to all four layers
RESEARCH QUESTIONS & HYPOTHESES
Core hypothesis: in an organizational setting, transcription and translation become significantly more accurate and natural when the system processes like a brain rather than translating isolated sentences.
RESEARCH QUESTIONS
How much more stable is speaker language identification when voiceprints are combined with conversational context, compared with audio-only detection?
How much does meeting-level working memory improve translation quality over sentence-by-sentence translation?
How much does topic-aware predictive processing reduce transcription errors on specialized terms?
Can multi-level translation (fast → sentence refinement → paragraph composition) balance latency and accuracy?
Does showing the AI's confidence (metacognitive transparency) increase user trust and adoption?
HYPOTHESES
Binding language with voiceprint plus context significantly reduces the language flap rate compared with baseline.
Translation using meeting-level working memory scores significantly higher than sentence-by-sentence translation.
Tuning the model with topic-based term banks significantly reduces specialized-term word error rate.
Users rate their understanding of multilingual meetings higher with Argus than without.
ROADMAP
Development follows the brain's own order, from perception to reasoning. Each phase builds on the last and must pass a measurable exit gate. Argus capabilities go live as each phase passes, without waiting for all five.
Phase 1 · Perception
Hearing speech in 5 languages
Gate 1 · WER/CER and DER meet targets vs. baseline
Phase 2 · Attention
Attention and language binding
Gate 2 · Language ID ≥ 95% · flap rate down ≥ 50%
Phase 3 · Understanding
Comprehension and memory
Gate 3 · Better quality than sentence-by-sentence · latency under 2 seconds
Phase 4 · Prediction
Prediction and organizational memory
Gate 4 · Domain-term WER down ≥ 30% · DPO approval
Phase 5 · Reasoning
Reasoning and metacognition
Cortex LLM v1.0, live in Argus
METRICS & EXIT GATES
Targets below are preliminary and will be set as absolute figures once baselines are measured in Phase 1.
| PHASE | METRIC | WHAT IT MEASURES | PRELIMINARY TARGET |
|---|---|---|---|
| 1 Perception | WER / CER per language | Transcription error rate (CER for Thai, Chinese, Japanese) | Thai ≥ 20% better than public baseline |
| 1 Perception | DER | Speaker diarization error rate | Below baseline on simulated meetings |
| 2 Binding & Attention | Language ID accuracy | Correctly identifies each speaker's language | ≥ 95% per speech segment |
| 2 Binding & Attention | Flap rate | Wrong language switches per speaker per hour | Down ≥ 50% from baseline |
| 3 Understanding & Memory | COMET / native-speaker raters | Meaning-level translation quality | Significantly better than sentence-by-sentence |
| 3 Understanding & Memory | First-display latency | Time from end of speech to first translation shown | < 2 seconds |
| 4 Prediction & Org Memory | Domain-term WER | Errors on organization-specific terms | Down ≥ 30% vs. general model |
| 5 Reasoning & Metacognition | Calibration error | How well shown confidence matches real accuracy | Below the threshold agreed in Phase 4 |
| 5 Reasoning & Metacognition | User understanding & trust | Post-meeting surveys in pilot organizations | Significantly better than control group |
Stability builds trust more than small accuracy gains.
A key lesson from designing Argus: a low flap rate, meaning language identification that doesn't keep switching, does more for how reliable the system feels than a slight rise in accuracy. That's why flap rate is the main gate for Phase 2.
METHODOLOGY
The research follows a Design Science Research approach combined with quantitative experiments: build the system, test it on standard datasets, evaluate it in real organizations, and repeat in every phase.
DATA
Public datasets
Baselines for speech recognition, diarization and translation across all 5 languages.
Simulated meetings
Multilingual recordings with designed scenarios such as language switching, overlapping speech and noise, with participant consent.
Thai business speech
Thai meeting audio mixed with English terms, with transcripts checked by experts.
Pilot organizations
Data from pilot partners under data-use agreements and PDPA.
EXPERIMENTS
E1 · Language binding
Audio-only detection vs. voiceprint + context
RQ1 · H1
E2 · Context memory
Sentence-by-sentence vs. with Meeting Context Buffer
RQ2 · H2
E3 · Specialized terms
General model vs. model with Topic-aware Prior
RQ3 · H3
E4 · Multi-level translation
One level vs. three levels, measuring latency and quality
RQ4
E5 · Field study
Multilingual meetings with vs. without Argus in pilot organizations
RQ5 · H4
ETHICS, PRIVACY & PDPA
Under Thailand's PDPA, voiceprints count as biometric data, so these principles apply from Phase 1 onward.
Consent first
Every participant in research data and pilot meetings gives written consent.
Meeting-only voiceprints
By default, voiceprints are deleted when the meeting ends and kept only with consent.
Audio stays on-premise
Audio and biometric data never leave the organization; only text goes to the cloud.
No outsider enrollment
Voices of people outside the meeting, such as a nearby table or booth, are never registered.
Transparency
Participants always know when recording and translation are on, and how identity and language are detected.
DPO at every change
The Data Protection Officer is consulted whenever data types or retention periods change.
No hidden evaluation
Meeting data isn't used to assess individual performance without a clear policy and consent.
OUTCOMES
Academic: a brain-inspired architecture for multi-speaker meeting translation, and new knowledge on speaker–language binding for Asian languages.
Resources: datasets and evaluation methods for code-switched Thai business meetings.
Product: Cortex LLM v1.0 as the core of Argus, connected to the Origami Platform.
Organizational: evidence that AI-supported multilingual communication improves meeting understanding and outcomes.
Closely related language pairs, such as Thai and Lao, may cause language identification errors.
Several people talking over each other remains a hard diarization problem.
Field results from a limited number of pilot organizations may not reflect every industry.
The brain metaphor is a design framework, not a biological simulation.
More languages, such as Vietnamese, Indonesian and other ASEAN languages.
Visual perception alongside audio, such as slides and documents shown on screen during meetings.
Connecting meeting signals to the Origami HR Skills Graph as a continuous evidence loop.
