TaskTalk Logo TaskTalk
Products/Document AI/OCR & Text Recognition
DOCUMENT INTELLIGENCE PIPELINE · STEP 02

Read every character. In every condition. Native Vietnamese OCR that preserves structure.

Transform scans, photos, and digital documents into crystal-clear digital text. TaskTalk is purpose-built for Vietnamese diacritics, complex multi-column layouts, and faded printouts. It doesn't just read words—it captures their exact position, hierarchy, and context.

99.3%
Vietnamese Tone Accuracy

Native Vietnamese neural model captures complex double diacritics and subtle tone marks.

< 1.2s
Processing Speed

Under 1.2s per page powered by high-throughput parallel asynchronous micro-workers.

98.8%
Dot-Matrix & Faded Recovery

Adaptive contrast recovery unscrambles thermal paper receipts and worn invoices.

0.00%
Hallucination Rate

Strict 2D bounding coordinate grounding guarantees every character traces back to raw pixels.

The First Bottleneck

Most document automation fails at the very first reading step.

Generic open-source OCR engines and international cloud APIs strip tone marks (dấu hỏi, ngã), misread stacked diacritics (ệ, ở, ử), and turn structured tables into an unreadable soup of plain text. When an OCR engine misreads a single decimal point on a 100-million VND repair estimate or drops the accent on a patient's name, the error compounds across every downstream step.

Brittle Text-Dump Approach Generic Open-Source & Cloud OCR

Scrambled Data & Stripped Diacritics

  • Stripped Accents: Converts "Nguyễn Văn Hưởng" into "Nguyen Van Huong", creating severe identity ambiguities in banking KYC and insurance audits.
  • Flattened Tabular Geometry: Flattens 4-column garage parts estimates into a single comma-separated string, hopelessly losing unit-price-to-part linkages.
  • Zero Spatial Tracing: Plain string output has no coordinate bboxes; downstream review teams cannot click to verify the source pixel in original scans.
  • Total Loss of ICR: Blindly ignores handwritten surveyor notes and adjustments scribbled into margins by on-site adjusters.
// Raw Generic Output: Flattened & Corrupted "Bao gia sua chua xe 30G-89241 Thay the can truoc so luong 1 don gia 3500000 Giam 10 phan tram Son lai canh cua..."
Spatial Neural Grounding tasktalk Native Layout OCR

100% Diacritic & Structural Fidelity

  • 100% Diacritic Preservation: Specialized multi-task Transformer preserves complex Vietnamese stacked tone marks (ệ, ở, ử, ẫ) with 99.3% accuracy.
  • Cell-Boundary Preservation: Preserves 2D bounding boxes and geometric table matrices with distinct column and row identifiers.
  • Bounding-Box Grounding: Every extracted word is tied to pixel coordinates [ymin, xmin, ymax, xmax] for instant audit validation.
  • Integrated Handwriting ICR: Detects and digitizes handwritten adjuster annotations, technician remarks, and corporate red stamps.
{"table_id": "estimate_rows", "cells": [{"row": 1, "desc": "Thay thế cản trước", "qty": 1, "unit_price": 3500000, "confidence": 0.998}]}
Live Inspection Workbench

Interactive OCR & Spatial Coordinate Inspector

Experience pixel-grounded Vietnamese text detection, multi-column bounding, and ICR recognition.

Display overlay: Polygons Active
DOCUMENT_P01_SCAN.TIF 300 DPI · 2480x3508
Deskew: +0.42° Contrast Normalized
GARA Ô TÔ TRƯỜNG PHÁT - HÀ NỘI Header 99.8%

Số 458 Minh Khai, Hai Bà Trưng, TP. Hà Nội · Hotline: 0912.839.xxx

BÁO GIÁ SỐ: BG-2025/0819 ID 99.6%

Ngày lập: 12/03/2025

CHỦ XE / ĐẠI DIỆN:
ĐẶNG HỮU NGUYỄN 99.4%
BIỂN SỐ KIỂM SOÁT:
30G-892.41 99.9%
Hạng mục sửa chữa & thay thế phụ tùng Table Grid Detected (4 Cols, 3 Rows)
STT Tên Phụ Tùng / Công Việc SL Thành Tiền (VND)
01 Thay thế cản trước chính hãng 01 3,500,000
02 Gò nắn tai xe bên phụ 01 850,000
03 Sơn sấy cản trước + tai xe 02 1,600,000
ICR · Surveyor Handwritten Note (95.7%)

"Khung xương bị dập nhẹ góc dưới, đã kiểm tra thước lái không lệch. Đồng ý duyệt mục 1 và 2."

GĐV: Trần Anh Tuấn (Ký ngày 12/03)
SEAL VERIFIED · 99.1%
DẤU
CÔNG TY TNHH GARA TRƯỜNG PHÁT Mã số thuế: 010892xxxx
Polygons Extracted: 48 tokens · 1 Table Matrix · 1 ICR Chunk
Latency: 940ms 0 Hallucinations
Normalized UTF-8 Entities & Confidence 100% Vietnamese UTF-8
VEHICLE_OWNER 99.4% conf
ĐẶNG HỮU NGUYỄN
bbox: [120, 45, 142, 280] · Vietnamese Double Diacritics OK
LICENSE_PLATE 99.9% conf
30G-892.41
bbox: [120, 310, 142, 420] · Format: VN Passenger Car
SURVEYOR_NOTE (ICR HANDWRITTEN) 95.7% conf
"Khung xương bị dập nhẹ góc dưới, đã kiểm tra thước lái không lệch. Đồng ý duyệt mục 1 và 2."
bbox: [410, 45, 490, 480] · Neural Stroke Synthesizer v3
RED_STAMP_DETECTION 99.1% conf
CON DẤU PHÁP NHÂN ĐẠT CHUẨN: CÔNG TY TNHH GARA TRƯỜNG PHÁT
Circle Bounding Radius: 54px · Overlap Angle: 0°
Synchronized Spatial Output JSON
{
  "document_type": "garage_repair_estimate",
  "engine_version": "v4.2-layout-aware",
  "entities": {
    "repair_total": 5950000,
    "currency": "VND",
    "diacritics_preserved": true,
    "table_rows_extracted": 3
  },
  "coordinates_anchored": true
}
Architectural Blueprint

How OCR & Text Recognition Works — 3 Distinct Sub-Stages

From degraded pixel ingest to geometry-grounded digital text ready for downstream schema validation.

STAGE 2.1

Enhance & Segment

Raw documents arrive skewed, poorly lit, or printed on thin thermal paper. Stage 2.1 performs non-destructive optical restoration before reading:

  • • Adaptive Contrast Enhancement: Unscrambles faded dot-matrix printouts and low-contrast carbon copy receipts.
  • • Micro-Deskew & Orientation: Automatically detects skew angles up to ±45° and corrects upside-down scans (0°, 90°, 180°, 270°).
  • • Layout Segmentation: Deconstructs documents into semantic visual zones: text blocks, table grids, seal zones, and stamps.
Output: Normalized 300 DPI Canvas + Zone Tree
STAGE 2.2

Recognize & Digitize

TaskTalk's neural reading engine executes token recognition trained extensively on enterprise Vietnamese administrative corpora:

  • • Zero Accent Stripping: Complete coverage for all Vietnamese diacritics (ă, â, đ, ê, ô, ơ, ư) and all 5 tone marks (huyền, sắc, hỏi, ngã, nặng).
  • • Printed OCR + Handwriting ICR: Jointly parses standard machine print alongside cursive surveyor notes.
  • • Character Confidence Scoring: Assigns explicit 0.00–1.00 probability scores to every recognized character and word.
Output: High-Confidence UTF-8 Stream
STAGE 2.3

Spatial Geometry & Mapping

Text without coordinates is useless for automated enterprise audits. Stage 2.3 anchors geometry to the underlying document:

  • • Preserve Natural Reading Order: Reconstructs accurate reading flows in complex multi-column forms and nested tables.
  • • Coordinate Grounding: Links each word to bounding coordinates, enabling click-to-source validation in Step 6 (Review & Approval).
  • • Dual-Layer Searchable PDF: Embeds an invisible text layer directly behind original scans for archival compliance.
Output: Georeferenced Spatial Document Schema
Core Differentiators

What Makes TaskTalk OCR Different

Engineered from ground zero for the reality of Vietnamese administrative, insurance, and medical paper chains.

Built for Real-World Vietnamese

Western models fail on Vietnamese because diacritics are stacked and tiny. TaskTalk distinguishes ở from ờ, ể from ễ, even on faded 58mm thermal receipts or degraded carbon paper. Handles mixed Vietnamese & English documents seamlessly.

Specialized Diacritic Attention Heads

Table & Column Geometry Preservation

Most OCR dumps words into a flat string, destroying row and column alignment. TaskTalk uses geometric computer vision to detect cell borders, row baselines, and vertical alignments—so tables stay structured as tables for downstream extraction.

Multi-Grid Spatial Graph Parsing

Stamp, Signature & Checkbox Detection

Vietnamese business documents rely on circular red stamps (con dấu đỏ) and ink signatures for legal validity. TaskTalk detects stamps, checks presence, classifies checkboxes (checked/unchecked), and flags un-stamped documents automatically.

Legal Seal & Form Field Verification
Empirical Field Validation

Recognition in Practice — Garage Repair Estimate

Document Type: 2-page garage loss assessment with faded print, grease smudges, and handwritten adjuster notes.

Evaluation Criteria Generic Cloud OCR TaskTalk Native OCR
Vietnamese Tone Accuracy
Stacked accents, hỏi/ngã distinction
87.4%
Strips subtle marks, misreads "ở" as "o", fails on legal names
99.3%
Native Vietnamese neural model captures complex double diacritics
Faded Dot-Matrix Numbers
Prices, parts codes, tax serials
Confuses '8' with '3' or 'B'
Creates massive VND calculation discrepancies downstream
98.8% Accurate
Adaptive contrast recovery unscrambles broken pin-printer dots
Table Layout Preservation
Item, Quantity, Unit Price, Total
Collapsed Text Stream
Merges parts description and numeric columns into flat strings
Preserved Grid Matrix
Row-and-column boundaries preserved with cell-level bounding boxes
Handwritten Surveyor Notes
Margin approvals & damage notes
Ignored or Garbage Symbols
Requires manual re-keying by human adjuster
Extracted via ICR & Tagged
Integrated handwriting model captures assessor remarks and tags them
Processing Speed
Per-page recognition latency
3.5s / page
High latency creates queue bottlenecks on claim spikes
< 1.2s / page
Parallel GPU inference optimized with TensorRT-LLM
Direct Pipeline Handoff

High-fidelity digitized text and spatial bounding boxes are immediately ready for Document Classification (Step 3) and Data Extraction & Validation (Step 4).

Explore Pipeline Flow →
End-to-End Orchestration

What Happens After Text Recognition

OCR provides the foundation. Follow the document through the complete 8-step TaskTalk intelligence pipeline.

STEP 01

Document Input

Multichannel capture via APIs, bulk SFTP, mobile SDK, and email listeners.

Status: Complete
YOU ARE HERE
STEP 02

OCR & Text Recognition

Native Vietnamese character reading, table layout geometry, and coordinate grounding.

Current Page
STEP 03

Document Classification

Automated categorization across 120+ Vietnamese claim and financial document types.

Downstream Step →
STEP 04

Data Extraction & Validation

Targeted field extraction, cross-field math checks, and tax authority database lookup.

Downstream Step →
STEP 05

AI-Powered Intelligence

Medical bill code matching, fraud pattern alerts, and historical duplicate detection.

Downstream Step →
STEP 06

Review & Approval

Smart routing with confidence-threshold exceptions for Human-in-the-Loop operators.

Downstream Step →
STEP 07

Processing Analytics

Operational STP dashboards, SLA monitors, auditor efficiency, and accuracy metrics.

Downstream Step →
STEP 08

Data Export & Integration

Direct webhooks, core banking CBS feeds, insurance claim engines, and cloud ERP.

Downstream Step →
Technical Clarifications

Frequently Asked Questions

Everything enterprise engineering teams need to know about our layout-aware OCR engine.

How does TaskTalk handle Vietnamese accents and diacritics?

TaskTalk's OCR model was trained from the ground up on Vietnamese text corpora, including historical documents, administrative forms, and degraded scans. Unlike Western OCR models that treat diacritics as noise or secondary accents, our neural network treats every Vietnamese character—including tone marks (huyền, sắc, hỏi, ngã, nặng) and modified vowels (ă, â, đ, ê, ô, ơ, ư)—as distinct first-class tokens.

Can TaskTalk read thermal paper receipts and faded scans?

Yes. Our preprocessing pipeline includes adaptive contrast enhancement, local binarization, and noise reduction specifically tuned for thermal paper receipts, faded dot-matrix printouts, and multi-generation photocopies. The system achieves 98.8% accuracy on low-contrast documents that standard OCR engines fail to process.

Does TaskTalk output plain text or structured layout?

TaskTalk outputs structured data with full spatial awareness. Every recognized word, line, and table cell includes its 2D coordinates (bounding box), reading order, confidence score, and hierarchical relationship to surrounding elements. We can also generate searchable, dual-layer PDFs that preserve the original visual document.

Can TaskTalk read handwritten text (ICR)?

Yes. TaskTalk includes an Intelligent Character Recognition (ICR) module trained on Vietnamese handwriting, including medical notes, insurance claim annotations, and filled form fields. While printed text achieves 99.3%+ accuracy, handwritten text achieves 95.7%+ on legible script, with confidence scoring flagging ambiguous characters for human review.

What is the average recognition speed per document?

Average processing time is under 1.2 seconds per page on standard cloud infrastructure. For high-volume enterprise deployments, TaskTalk supports parallel batch processing capable of handling 50,000+ pages per hour with auto-scaling GPU workers.

Does TaskTalk work on photos taken by mobile phones?

Yes. Our image preprocessing pipeline automatically corrects perspective distortion, shadows, uneven lighting, blur, and skew commonly found in smartphone captures. Mobile photos go through automatic border detection and deskewing before text recognition begins.

Zero Integration Friction · REST API & On-Premise

Ready to stop dealing with broken accents and scrambled tables?

Upload a sample document and see how TaskTalk's native Vietnamese OCR digitizes text, tables, and diacritics with unmatched precision.

Test Your Documents

Enterprise direct line: hello@tasktalk.com · Dedicated Sandbox API keys issued in 15 minutes.