AISTATEweb Community (3.7.2 beta)

English | Polski | 한국어 | Español | Français | 中文 | Українська | Deutsch

---

AISTATEweb Community is a web-based tool for audio transcription, speaker diarization, translation, AI-powered analysis, and structured reporting — fully offline, running on local hardware.

Feedback / Support

If you have any issues, suggestions, or feature requests, please contact me at: pawlict@proton.me

---

🚀 Main Functionalities

🎙️ Speech Processing

• Automatic speech recognition (ASR) using Whisper, WhisperX, and NVIDIA NeMo
• Support for multilingual audio (PL / EN / UA / RU / BY and more)
• Offline and local model execution (no cloud dependency)
• High-quality transcription optimized for long recordings

🧩 Speaker Diarization

• Advanced speaker diarization using pyannote and NeMo Diarization
• Automatic speaker detection and segmentation
• Support for multi-speaker conversations (meetings, interviews, calls)
• Configurable diarization engines and models

🌍 Multilingual Translation

• Neural machine translation powered by NLLB-200
• Fully offline translation pipeline
• Flexible source and target language selection
• Designed for OSINT and multilingual analysis workflows

🧠 Intelligence & Analysis

• AI-assisted content analysis using local LLM models
• Transformation of raw speech and text into structured insights
• Support for analytical reports and intelligence-oriented workflows

📱 GSM / BTS Analysis

• Import and analysis of GSM billing data (CSV, XLSX, PDF)
• Interactive map visualization of BTS locations (Leaflet + OpenStreetMap)
• Offline maps support via MBTiles (raster PNG/JPG/WebP + vector PBF via MapLibre GL)
• Multiple map views: all points, path, clusters, trips, BTS coverage, heatmap, timeline
• Area selection (circle / rectangle) for spatial queries
• Overlay layers: military bases, civilian airports, diplomatic posts (built-in data)
• KML/KMZ import — custom overlay layers from Google Earth and other GIS tools
• Map screenshots with watermark (online & offline maps + all overlay layers)
• Contact graph, activity heatmap, top contacts analysis
• Timeline player with month/day animation

💰 AML — Financial Analysis

• Anti-Money Laundering analysis pipeline for bank statements
• Automatic bank detection and PDF parsing for Polish banks: PKO BP, ING, mBank, Pekao SA, Santander, Millennium, Revolut (+ generic fallback)
• MT940 (SWIFT) statement format support
• Transaction normalization, rule-based classification, and risk scoring
• Anomaly detection: statistical baseline + ML-based (Isolation Forest)
• Graph analysis — counterparty network visualization
• Cross-account analysis for multi-account investigations
• Entity resolution and counterparty memory (persistent labels/notes)
• Spending analysis, behavioral patterns, merchant categorization
• LLM-assisted analysis (prompt builder for Ollama models)
• HTML report generation with charts
• Data anonymization profiles for safe sharing

🔗 Crypto — Blockchain Transaction Analysis (experimental)

• Offline analysis of BTC and ETH cryptocurrency transactions
• Import from WalletExplorer.com CSV and multiple exchange formats (Binance, Etherscan, Kraken, Coinbase, and more)
• Automatic format detection from CSV column signatures
• Risk scoring with pattern detection: peel chain, dust attack, round-trip, smurfing
• OFAC sanctioned address database and known DeFi contract lookup
• Interactive transaction flow graph (Cytoscape.js)
• Charts: balance timeline, monthly volume, daily activity, counterparty ranking (Chart.js)
• LLM-assisted narrative analysis via Ollama
• This module is currently in early testing phase — features and data formats may change

⚙️ GPU & Resource Management

• Integrated GPU Resource Manager
• Automatic task scheduling and prioritization (ASR, diarization, analysis)
• Safe execution of concurrent tasks without GPU overload
• CPU fallback when GPU resources are unavailable

📂 Project-Based Workflow

• Project-oriented data organization
• Persistent storage of audio, transcripts, translations, and analyses
• Reproducible analytical workflows
• Separation of user data and system processes

📄 Reporting & Export

• Export results to TXT, HTML, DOC, and PDF
• Structured reports combining transcription, diarization, and analysis
• AML financial reports with charts and risk indicators
• Ready-to-use outputs for research, documentation, and investigations

🌐 Web-Based Interface

• Modern web UI (AISTATEweb)
• Real-time task status and logs
• Multi-language interface (PL / EN)
• Designed for both standalone and multi-user environments (soon)

---

Requirements

System (Linux)

Install base packages (example): sudo apt update -y sudo apt install -y python3 python3-venv python3-pip git

Python

Recommended: Python 3.11+.

---

pyannote / Hugging Face (required for diarization)

Diarization uses pyannote.audio pipelines hosted on the Hugging Face Hub. Some pyannote models are gated, which means you must:

• have a Hugging Face account,
• accept the user conditions on the model pages,
• generate a READ access token and provide it to the app.

Step-by-step (token + permissions)

1. Create / sign in to your Hugging Face account.
2. Open the required pyannote model pages and click "Agree / Accept" (user conditions). Typical models you may need to accept (depending on version):
   • pyannote/segmentation (or pyannote/segmentation-3.0)
   • pyannote/speaker-diarization (or pyannote/speaker-diarization-3.1)
3. Go to your Hugging Face Settings → Access Tokens and create a new token with role READ.
4. Paste the token into AISTATE Web settings (or provide it as an environment variable — depending on your setup).

---

Installation (Linux)

sudo apt update
sudo apt install -y ffmpeg
curl -fsSL https://ollama.com/install.sh | sh

mkdir -p ~/projects
cd ~/projects
git clone https://github.com/pawlict/AISTATEweb.git
cd AISTATEweb

python3 -m venv .venv
source .venv/bin/activate

python -m pip install --upgrade pip wheel setuptools
pip install -r requirements.txt

---

Run

python3 AISTATEweb.py

Example (uvicorn): python -m uvicorn webapp.server:app --host 0.0.0.0 --port 8000

Open in browser: http://127.0.0.1:8000

---

AISTATEweb — Windows (WSL2 + NVIDIA GPU) Setup

Important: In WSL2 the NVIDIA driver is installed on Windows, not inside Linux. Do not install nvidia-driver-* packages inside the WSL distro.

---

1. Windows side

1. Enable WSL2 (PowerShell: wsl --install or Windows Features).
2. Install the latest NVIDIA Windows driver (Game Ready / Studio) — this provides GPU support inside WSL2.
3. Update WSL and restart:

   wsl --update
   wsl --shutdown
   

2. Inside WSL (Ubuntu recommended)

sudo apt update && sudo apt upgrade -y
sudo apt install -y git python3 python3-venv python3-pip ffmpeg

Verify GPU is visible:

nvidia-smi

3. Install AISTATEweb

mkdir -p ~/projects && cd ~/projects
git clone https://github.com/pawlict/AISTATEweb.git
cd AISTATEweb

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip setuptools wheel

# PyTorch with CUDA (example: cu128)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

pip install -r requirements.txt

Verify GPU access:

python -c "import torch; print('CUDA:', torch.cuda.is_available(), torch.cuda.get_device_name(0) if torch.cuda.is_available() else '')"

4. Run

python3 AISTATEweb.py

Open in browser: http://127.0.0.1:8000

Troubleshooting

If nvidia-smi doesn't work inside WSL, make sure you did not install Linux NVIDIA packages. Remove them if present:

sudo apt purge -y 'nvidia-*' 'libnvidia-*' && sudo apt autoremove --purge -y

---

References

• NVIDIA: CUDA on WSL User Guide
• Microsoft: Install WSL
• PyTorch: Get Started
• pyannote.audio (Hugging Face)
• Whisper (OpenAI)
• NLLB-200 (Meta)
• Ollama

---

"This project is MIT licensed (AS IS). Third-party components are licensed separately — see THIRD_PARTY_NOTICES.md."

beta 3.7.2

• Analyst panel — new sidebar panel replacing notes sidebar in transcription and diarization pages
• Block notes with tags — notes can now have colored tags, shown as left border on segments
• Revolut crypto PDF — parser for Revolut cryptocurrency statements, integrated with AML pipeline
• Token database (TOP 200) — known/unknown token classification for crypto analysis
• Improved reports — DOCX/HTML reports with charts, watermarks, dynamic conclusions, section descriptions
• Fixed translation stuck at 5% (auto-detect model cache)
• Fixed translation report losing formatting (newlines collapsed)
• Fixed stale transcription/diarization results on new audio upload
• No-cache middleware for static JS/CSS files

beta 3.7.1

• Cryptocurrency Analysis — Binance — extended analysis of Binance exchange data
• User behavior profiling (10 patterns: HODLer, Scalper, Day Trader, Swing Trader, Staker, Whale, Institutional, Alpha Hunter, Meme Trader, Bagholder)
• 18 forensic analysis cards: internal counterparties, Pay C2C, on-chain addresses, pass-through flows, privacy coins, access logs, payment cards + NEW: temporal analysis, token conversion chains, structuring/smurfing detection, wash trading, fiat on/off ramp, P2P analysis, deposit-to-withdrawal velocity, fee analysis, blockchain network analysis, extended security (VPN/proxy)
• All record limits removed — full data with scrollable tables
• Reports download as files (HTML, TXT, DOCX)

beta 3.7

• Crypto Analysis (experimental) — offline blockchain transaction analysis module (BTC/ETH), CSV import (WalletExplorer + 16 exchange formats), risk scoring, pattern detection, flow graph, Chart.js charts, LLM narrative — currently in deep testing phase
• Auto-detect source language on file upload and text paste (translation module)
• Multi-language export (all translated languages at once)
• Fixed DOCX export filenames (underscores issue)
• Fixed MMS TTS waveform synthesis error
• Fixed Korean language missing from translation results

beta 3.6

• GSM / BTS Analysis — full GSM billing analysis module with interactive maps, timeline, clusters, trips, heatmap, contact graph
• AML Financial Analysis — anti-money laundering pipeline: PDF parsing (7 Polish banks + MT940), rule-based + ML anomaly detection, graph analysis, risk scoring, LLM-assisted reports
• Map overlays — military bases, airports, diplomatic posts + custom KML/KMZ import
• Offline maps — MBTiles support (raster + PBF vector via MapLibre GL)
• Map screenshots — full map capture including all tile layers, overlays, and KML markers
• Fixed KML/KMZ parser (ElementTree falsy element bug)
• Fixed MapLibre GL canvas screenshot (preserveDrawingBuffer)
• Fixed info page language switching

beta 3.5.1/3

• Fixed project saving/assignment.
• Improved parser for ING banking

beta 3.5.0 (SQLite)

• JSON -> SQLite migration

beta 3.4.0

• Added multiuser

beta 3.2.3 (translation update)

• Added Translation module
• Added NLLB Settings page
• Added the ability to change task priorities
• Added Chat LLM
• Background sound analysis (experimental)

beta 3.0 - 3.1

• LLM Ollama modules for data analysis introduced
• GPU Assignment / Scheduling (Update)

This update introduces a GPU Resource Manager concept in the UI and internal flow to reduce the risk of overlapping GPU-heavy workloads (e.g., running diarization + transcription + LLM analysis at the same time).

What problem this solves

When multiple GPU tasks start concurrently, it can lead to:

• sudden VRAM exhaustion (OOM),
• driver resets / CUDA errors,
• extremely slow processing due to contention,
• unstable behavior when multiple users trigger jobs at the same time.

Backwards compatibility

• No changes in the functional layout of existing tabs.
• Only GPU admission/coordination and admin labeling were updated.

beta 2.1 -2.2

• Change of block editing methodology
• This update focuses on improving observability and usability of application logs.
• Fix: Logging overhaul (Whisper + pyannote) + Export to file