57 lines
3.3 KiB
JSON
57 lines
3.3 KiB
JSON
{
|
|
"title": "Classification Guide for Human Language Technology (HLT) Models (May 18, 2011)",
|
|
"type": "REPORTED",
|
|
"description": "This document is an NSA administrative policy document that establishes the rules for what information regarding speech analytics remains public versus what must be kept secret. It explicitly confirms that the NSA possesses automated tools for gender, language, dialect, and speaker recognition, as well as speech-to-text and phonetic tokenization. While the existence of these capabilities is unclassified, the specific data used to \"train\" the models and the actual identity of targeted speakers are protected at the SECRET level.",
|
|
"long_description": "This guide serves as the legal framework for \"Human Language Technology Research\" (R66E), defining how the NSA manages the transition of raw SIGINT into machine-learning models. A key admission in the \"General\" section (A.2) is that these HLT models are built by \"aggregating statistics derived from SIGINT collection.\" This means the NSA uses the vast quantities of voice traffic it intercepts globally to refine its algorithms, effectively using every captured conversation to train its computers to better recognize future targets.\n\nThe guide makes a critical distinction between Acoustic Models and Language Models:\n\n* Acoustic Models (A.9): These handle the physical sounds of speech (phonetic tokenization). The guide notes that while the models are trained on classified messages, the original audio cannot be \"recovered\" from the mathematical model itself. Therefore, the models are capped at SECRET//REL TO USA, FVEY to protect the source of the data.\n\n* Language Models (A.10): These handle the \"what\" of the conversation (vocabulary and grammar). Because these models can reveal the specific content or entities being targeted, their classification is much stricter, matching the highest classification of the targeted entities or messages used in the training set.",
|
|
"tags": [
|
|
"linguistics",
|
|
"speech-to-text",
|
|
"machine translation",
|
|
"NLP",
|
|
"natural language processing",
|
|
"phonetic search",
|
|
"gender",
|
|
"speaker identification",
|
|
"dialects",
|
|
"voice analytics",
|
|
"transcription",
|
|
"acoustic modeling",
|
|
"pattern recognition",
|
|
"neural networks"
|
|
],
|
|
"document": {
|
|
"creator": "NSA",
|
|
"date_created": "2011-05-18",
|
|
"classification": [
|
|
"UNCLASSIFIED"
|
|
],
|
|
"dissemination": [
|
|
"FOUO"
|
|
],
|
|
"banner": "UNCLASSIFIED// FOR OFFICIAL USE ONLY",
|
|
"genre": "guide",
|
|
"programs": [
|
|
"HLT"
|
|
],
|
|
"countries": [],
|
|
"people": []
|
|
},
|
|
"publication": [
|
|
{
|
|
"publisher": "The Intercept",
|
|
"date_published": "2015-05-05",
|
|
"title": "How the NSA Converts Spoken Words Into Searchable Text",
|
|
"link": "https://theintercept.com/2015/05/05/nsa-speech-recognition-snowden-searchable-text/",
|
|
"reporters": [
|
|
"Froomkin, Dan"
|
|
],
|
|
"contributors": [
|
|
"Fishman, Andrew"
|
|
],
|
|
"doc_source": "https://theintercept.com/document/2015/05/05/classification-guide-human-language-technology-hlt-models/",
|
|
"doc_url": "https://s3.documentcloud.org/documents/2072015/human-language-technology.pdf"
|
|
}
|
|
],
|
|
"pdf_sha256": "f4d5d14adf05036715c2d5036cf9ff000534381e9904e394ff21e19c3085d66d",
|
|
"pages": 4
|
|
}
|