M3: Arabic meeting transcription
Evaluation report ·1.62% character error rate on real multi-dialect Arabic meetings with human-verified transcripts, the lowest of seven systems scored under the same rules. Google chirp_3 was next at 2.26%.
Company
The research is the moat.
Read it yourself.
Papers written from production systems, inventions nobody else has built, and benchmarks we measured instead of promised.
Lisan Research
Evaluation reports.
Evaluations we designed and ran ourselves, published with the method, the competitors' versions and the limits.
L3: Arabic language checking
Evaluation report ·80.29% edit F0.5 on a linguist-reviewed benchmark of real Arabic sentences, against 44.68% for Claude Fable 5.1, the strongest of 26 large language model configurations.
Why we publish
Deep tech you can read,
not just believe.
Behind the products sits over a decade of language engineering: 19.5 million lines of internally developed software, 32 AI components, and a data operation that never stops. The papers below document how the hardest parts actually work, and the innovations list shows the parts nobody else has built.
The core inventions
Relay Architecture and
Morphological Compression.
Most language AI throws compute at text one opaque layer at a time. Lisan's core stack does the opposite: compress the language using its own morphology, then process it through explicit, explainable spaces. The result is energy-efficient, sovereign AI that runs where transformers cannot, from your data center down to edge devices.
وبالكتاب
وبالكتابجذر: ك·ت·ب
99.97%of compute saved
Arabic words decompose into roots and morphemes, so the representation shrinks far below traditional tokenization, losslessly. vs traditional tokenization in transformer models
Morphological
Semantic
Zipf
Latent
The compressed stream crosses explicit, explainable spaces instead of one black box.
4.7× fasterExplainable by designRuns on edge devices
Papers
Ten papers from
inside the platform.
Written from production systems, not lab prototypes. Each one documents a problem the platform had to solve for real customers, prepared for peer-reviewed publication.
A Hybrid Rule-Based and Neural Architecture for Production Arabic NLP
How the Lisan Engine pairs a ~10M-line rule-based Java engine with a neural model server over gRPC, routing each request to the engine that does it best: deterministic sub-second grammar checking on one side, transformer-scale paraphrasing and summarization on the other.
Perfect Speaker Diarization from Per-Participant Audio Streams
Traditional diarization guesses who spoke when and still gets 5-15% of it wrong. By capturing a separate audio stream per meeting participant, Lisan makes speaker identity deterministic: the error rate of guessing drops out of the equation entirely.
Arabic Morphological Analysis at Scale
A production morphological analyzer that decomposes any Arabic word into root, pattern, and stem, handling the clitics and missing diacritics that give a single token dozens of valid readings, at speeds that keep grammar checking interactive.
Arabic OCR with Language-Model Correction
Arabic's cursive script and contextual letter shapes break conventional OCR. OCR PLUS combines multiple recognition engines and corrects their output with an n-gram Viterbi decoder, so the language model fixes what the vision models miss.
Layout-Preserving Document Translation
Translating a document is easy; giving it back with the layout intact is not. This work covers TranslateX's format-aware pipeline for DOCX, XLSX, PPTX, PDF, and images, including the bidirectional-text handling Arabic demands.
Adaptive Segment Merging for Meeting Transcription
Raw ASR output is fragmented and full of platform artifacts. Eleven specialized merging heuristics, tuned per meeting platform, turn it into transcripts that read the way the meeting actually sounded.
A 162-Pipeline Data Architecture for Continuous NLP Improvement
The Lisan data layer runs 162 specialized pipelines: scrapers, corpus processors, n-gram calculators, entity extractors, and dictionary builders in a continuous feedback loop that makes the production engine better every week.
Keyboard-Aware Spell Correction
Spell-check models train best on realistic mistakes. By modeling which keys sit next to each other on a real keyboard, synthetic typos match the errors humans actually make, and the corrector learns the right lessons.
Deterministic Text Quality Assessment
LLM-based scoring gives a different answer every run. This system scores text with deterministic linguistic metrics instead, so the same paragraph always gets the same score, and the riskiest paragraphs always surface first.
A Multi-Platform Meeting Bot Architecture
One bot controller, four meeting platforms: Zoom, Microsoft Teams, Google Meet, and Webex, each behind an interchangeable adapter. The architecture that lets MeetriX join any meeting and leave with the minutes.
Measured, not promised
M3 records the lowest character error rate of seven systems
on the same human-verified meetings.
Lower is better. Character error rate (CER) after defined lexical normalisation. Lisan evaluation, September 2026.
Innovations
Built here, nowhere else.
A selection from the platform's invention portfolio: from the language core that reads Arabic the way it is actually built, to files that carry their own identity.
The language core
6
Jader
Lisan's morphological analyzer decodes the full structure of any Arabic word, surfacing every valid reading despite the missing short vowels that defeat most Arabic NLP.
Lisan Neuro-Statistical Model
Statistical and neural models fused into one hybrid architecture, removing the usual trade-off between speed, explainability, and semantic depth.
Canonical Form
An error-tolerant representation of Arabic words that lets misspelled and variant forms map to the same identity, simplifying search, matching, and correction.
Arabic Synthesizer
Generates correct Arabic words from stems and grammatical features, handling tense, gender, and affixes with full morphological fidelity.
Advanced Tashkeel
Deep-learning diacritization that combines morphological structure with neural networks, then ensembles four models by majority vote for accuracy no single model reaches.
Wildcard N-Gram Trie
A wildcard-enabled trie data structure engineered for long n-gram matching in morphologically rich languages, retrieving complex patterns at production speed.
Documents & files
6
FilePersona
Self-describing files that carry their own icon, color, and visual identity in embedded metadata, rendered live by the operating system shell.
VAX: Verifiable Autonomous Documents
Steganography and cryptography embed a public key inside any document without changing its size or appearance, making documents verifiable and programmable.
LODEX
A family of layout and document extraction models that read structure, not just text, from PDFs and images.
OCR+
A document intelligence engine that extracts text, layout, and structured data from images and PDFs, with Lisan's language models correcting what the vision models miss.
Doc2Text
Extracts text other libraries cannot reach: headers, footers, text boxes, WordArt, and non-standard Office files included.
Screen Text Extractor
Lifts text from any object in any program on the screen, ready to check, translate, or reuse.
Speech & meetings
3
LIPAS
A multi-stage Arabic speech recognition pipeline: preprocessing, neural decoding, and post-processing tuned together as one system.
Speaker Identity Mapping
Maps diarized voices to real participant names from meeting metadata, so transcripts say who spoke, not "Speaker 2".
Per-Speaker Meeting Report
Meeting summaries organized by speaker, with each person's action items, deadlines, and priorities extracted automatically.
Inside the products
5
Text2Chart
Turns written text into editable, brand-styled charts directly inside PowerPoint and Word.
Agentic Add-In Document Creation
An embedded AI agent inside Word and PowerPoint that drafts documents, edits content, and executes actions in place.
InText Queries & InText Search
A miniature query language that runs data filtering and lookups inside any text area, no spreadsheet detour required.
Automatic Layout Detection
Detects when you are typing with the wrong keyboard language and converts the text automatically.
Augmenters
Hover information boxes that recognize named entities in text and surface who or what they are, in place.
The data operation
Models are only as good
as what feeds them.
The Lisan Lexicon
A curated lexicon of modern standard Arabic built by intersecting expert dictionaries with live corpora: 681,771 expert-verified words, plus 14,300 modern words validated by hand.
LiStudio annotation
Our in-house annotation studio combines active learning and reinforcement learning with a team of 32 annotators, so every correction a human makes teaches the models.
A 32-dialect audio corpus
Speech data spanning modern standard Arabic and 32 dialects across 17 countries. It is why the transcription holds up when the meeting switches from Cairo to Riyadh mid-sentence.
27Btokens crawled into the training corpora
42Mwords in the morphologically annotated lexicon
57M+sentences behind the diacritization models
162data pipelines in continuous operation
Want the details?
For research collaborations, benchmark methodology, or a technical deep dive under NDA, talk to the team that built it.
Frequently asked questions
Are the papers peer-reviewed?
They are written from production systems and prepared for submission to peer-reviewed venues. The engineering they describe is running in products today; the publication process is underway.
Can we see the benchmark methodology?
Yes, it is published. The M3 evaluation report states the protocol (same audio, same human-verified reference, same rules for all seven systems), the full scoring rules for CER after normalisation, the configurations of every compared system, five transcript examples, an evidence map with the status of every claim, and the limitations. Read it at the M3 report. For a pilot on your own meetings, talk to us.
Start with one product.
Keep the whole platform.
Open a free account and solve today's problem in the next ten minutes. When you are ready for more, six flagships and a 20+ app workspace are already on your account. Or talk to us and we will scope it with you.
Free to start, no card · Your data exportable, always · Trusted by more than 24 government entities