Skip to content

A tool for researchers and Law Enforcement Agencies (LEAs) to process intercepted conversations in ongoing investigations

Notifications You must be signed in to change notification settings

idiap/autocrime

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 

Repository files navigation

Welcome to autocrime, a tool for researchers and Law Enforcement Agencies (LEAs) to process intercepted conversations in ongoing investigations.
More details about the platform (access, how to use, etc.) can be found on https://www.roxanne-euproject.org/platform. Briefly, the tool expects set of intercepted audio files as an input, and will run technologies (a set of technologies given below in table) to produce a JSON file as an output. This README summarizes performance of these automatic machine learning technologies when evaluated on ROXSD data (https://publications.idiap.ch/publications/show/5321) as well as on other (publicly available) datasets.

This README is aligned with current release of Autocrime (version: 1.4.0)

1. Supported Technologies

  • A summary of the integrated technologies to-date is presented in this table:

1.1 Speech/Audio Processing

Technology Partner(s) Method Performance (Other datasets) Performance ( ROXSD )
Closed set Speaker ID BUT & Phonexia ResNet architecture 99.95% speaker accuracy (subset of ROXSD) 96.49% accuracy (plda), 96.92% accuracy (cosine)
Open set Speaker ID BUT Same as SID 90% accuracy (subset of ROXSD) 93.43% (with ROXSD threshold 30)
Gender ID Idiap x-vector based 90% accuracy (subset of ROXSD) 75.08% accuracy
ASR (English) Idiap XLSR-LFMMI + 4gram LM 23% WER on SLURP 28.4% WER
ASR (German) Idiap XLSR-LFMMI + 4gram LM 11.1% on TUDA-v4 35.9% WER
ASR (Greek) Idiap XLSR-LFMMI + 4gram LM 12.72% WER TED talk Greek -
ASR (Spanish) Idiap XLSR-LFMMI + 4gram LM 7.8% on Voxpopuli -
ASR (Arabic) Idiap XLSR-LFMMI + 4gram LM 14.9% WER on GALE arabic -
ASR (Dutch) Idiap XLSR-LFMMI + 4gram LM 5.9% WER CGN -
ASR (Lithuanian) Idiap XLSR-LFMMI + 4gram LM 20.0% WER on MATERIAL (CS,NTB) -
Word boosting (English) Idiap Lattice Rescoring - 28.46% WER [impact of boosted ASR is seen on NER and Mention network]
Word boosting (German) Idiap Lattice Rescoring - 35.95% WER [impact of boosted ASR is seen on NER and Mention network]
Speaker Diarization BUT & Phonexia Energy-based VAD + VBx DER 5.91% on CALLHOME 14.8% DER

1.2 Natural Language Processing

Technology Partner(s) Method Performance (Other datasets) Performance ( ROXSD)
Topic Detection (English) Idiap Cross-Encoder (distilroberta) 76.72% accuracy(ground-truth), 75% accuracy (subset of ROXSD) (ASR) 28.66% (ground-truth), 21.34% (ASR) accuracy
NER (English) USAAR RoBERTa-large 79% F1-Score (subset of ROXSD) 82.8% F1 (ground-truth), 39.7% F1 (ASR), 43.24% F1 (Boosted ASR)
NER (German) USAAR RoBERTa-large - 70.1% F1 (ground-truth), 27.9% F1 (ASR), 30.57% F1 (Boosted ASR)
Mention network USAAR Custom co-reference analysis module 82% accuracy 74.81% accuracy (ground-truth), 71.62% accuracy (ASR), 75.29% accuracy (Boosted ASR)

1.3 Network Analysis

Technology Partner(s) Method Performance (Other datasets) Performance ( ROXSD)
Community Detection LUH (methods) & UCSC (assessments) Greedy Modularity F1-score 75% (subset of ROXSD) for communities based on speaker languages F1-score 32.8
Social Influence Analysis LUH (methods) & UCSC (assessments) Pagerank Speaker Accuracy (using Degree Centrality): 93.5% in ROXANNE Speaker Network and 80.2% in Telephone Network (subset of ROXSD) 95% in ROXANNE network, 79.5% in Telephone Network
Link Prediction LUH (methods) & UCSC (assessments) Jaccard Coefficient 67.2% accuracy (Top-5) (Burglary dataset) 58.82% accuracy (Top-5)
Outlier Detection LUH (methods) & UCSC (assessments) Pagerank & Threshold=0.3 - 100% accuracy with Threshold less than or equal 0.3
Cross-Network Node Matching LUH (methods) & UCSC (assessments) Node2Vec and DeepLink - Node Matching:
top-1 accuracy: 75%
on ROXSD with 60% nodes as training data

1.4 Visual Analytics

Technology Partner(s) Method Performance (Other datasets) Performance ( ROXSD)
Face detection and similarity match AIRBUS RetinaFace + ArcFace 98.06% on MegaFace, 99.8% on LFW 98% recall and 100% precision on ROXSD-videos
Scene characterization and similarity match AIRBUS ResNet + ArcFace N/A 70% recall and 86% precision on ROXSD-videos

2. Supported Languages

The list of current supported languages is presented below:

Language ASR NER Topic
English
German x
Arabic x x
Spanish x x
Greek x x
Dutch x x
Lithuanian x x

About

A tool for researchers and Law Enforcement Agencies (LEAs) to process intercepted conversations in ongoing investigations

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published