first commit

This commit is contained in:
libroot 2026-11-19 04:39:11 +02:00
commit 0d90363116
Signed by: libroot
GPG key ID: 2AC68C029EBE9ACE
6510 changed files with 679146 additions and 0 deletions

View file

@ -0,0 +1,13 @@
Speaker ID: since the last VFUG in July we have deployed:
- Two systems for OPI~MENA: one against an Iraqi political target and one against a
Saudi leadership target.
- One system for OPI~RCIT targeting a Gazprom official.
- A 5 way multi-speaker system for OPI~LANG/OPI~SC against Afghan counter
narcotics targets
We have now cleared our order book for speaker id systems and so look forward to a
well deserved, quiet, peaceful and reflective Autumn. Seriously though, if you believe
we can help you with identifying your target of interest amongst the deluge of traffic
that you have to wade through, feel free to approach us and we will happily discuss
your requirements and hopefully offer a swift and accurate solution.

View file

@ -0,0 +1,87 @@
{
"title": "Voice/Fax User Group Minutes of September 2008 meeting",
"type": "REPORTED",
"description": "Page 3 of 7 of the minutes of GCHQ's Voice/Fax User Group (VFUG), September 2008: plans for data-quality testing, a request that linguists flag \"Pick Your Own Number\" calls in the B3M voice database, and the B14/MCS team's update on the Diariser and on speaker and language identification. The Intercept's PDF shows only the speaker ID section; the rest of the page is in the file (see the archive comment).",
"long_description": "The page opens with the next phase of User Confidence (data quality) testing, part of the next \"Stepping Stone\" and due to start on 9 October 2008, to be carried out by the newly formed Content PUT. A suggestion from the LISTEN 08 briefing, agreed at the meeting, asks linguists to record \"Pick Your Own Number\" (PYON) cases in B3M with a 'techref PYON' comment: calls where a speaker on a UK-code number turns out to be based in another country.\n \nItem 4 is the B14/MCS update on voice processing. Version 5.0 of the Diariser was deployed over the summer; its main change is a rewritten modem arbiter that better identifies V-series modems so that OPC~CAP can demodulate them. Since July, new speaker ID systems were deployed for OPI~MENA (against an Iraqi political target and a Saudi leadership target), for OPI~RCIT (against a Gazprom official) and for OPI~LANG/OPI~SC (a five-way multi-speaker system against Afghan counter-narcotics targets). Language ID had no new deployments: models were being improved with non-SIGINT language data from speech-to-text corpora, which the team notes is no substitute for linguist-verified SIGINT data, and identifying one target language among unidentified others is said to produce false positives.\n",
"tags": [
"speaker identification",
"voice analytics",
"target identification",
"multi-speaker systems",
"leadership targets",
"counternarcotics",
"energy sector",
"Gazprom",
"VFUG",
"OPI-MENA",
"OPI-RCIT",
"OPI-LANG",
"OPI-SC",
"traffic volume",
"diarisation",
"modem identification",
"fax",
"OPC-CAP",
"language identification",
"speech to text",
"training data",
"false positives",
"data quality testing",
"User Confidence Testing",
"Stepping Stone",
"Content PUT",
"LISTEN 08",
"Pick Your Own Number",
"PYON",
"B14",
"MCS",
"minutes"
],
"document": {
"creator": "GCHQ",
"date_created": "2008-09",
"classification": [
"TOP SECRET",
"STRAP 1"
],
"dissemination": [],
"banner": "TOP SECRET STRAP1",
"genre": "report",
"programs": [
"DIARISER",
"B3M"
],
"countries": [
"Iraq",
"Saudi Arabia",
"Russia",
"Afghanistan"
],
"people": []
},
"publication": [
{
"publisher": "The Intercept",
"date_published": "2018-01-19",
"title": "Finding Your Voice: Forget About Siri and Alexa — When It Comes to Voice Identification, the “NSA Reigns Supreme”",
"link": "https://theintercept.com/2018/01/19/voice-recognition-technology-nsa/",
"reporters": [
"Kofman, Ava"
],
"contributors": [],
"doc_url": "https://s3.documentcloud.org/documents/4352015/Pages-From-VFUG-minutes-Sept-2008.pdf"
}
],
"aka_id": [],
"archive_comment": "The Intercept's PDF shows only an excerpt (the Speaker ID section) through a soft crop. The file itself contains the whole of page 3 of 7, including GCHQ staff names; the PDF is held as published. The page image and transcript show the visible excerpt only. The cropped-out content was described by Libroot ('Going Through Snowden Documents, Part 9' 2026-10-11 https://libroot.org/posts/going-through-snowden-documents-part-9/), with the names redacted.",
"pdf_sha256": "b32baeff6d9d1637a841e835e35f275281c67902e1a8f5ec9bde1f27ef614fbd",
"pages": 1,
"versions": {
"vfug-libroot": {
"pdf_sha256": "80b5663fd3460f259f6327800ad78ee7c9f4e3b7345e6ca0556243386af4751b",
"changed_pages": [
1
]
}
}
}

View file

@ -0,0 +1,56 @@
TOP SECRET STRAP1
[REDACTED] reported: the next phase will see User Confidence Testing (Data Quality
Testing) performed as part of the next Stepping Stone. This is due to start on 9
October. The Content PUT has now been formed and will carry out the User
Confidence Testing.
[REDACTED] suggested at [REDACTED]'s LISTEN 08 briefing that linguists record occurrences
of Pick Your Own Number (PYON) in B3M using the 'techref' method in the
proc_cmnt_text field. So, if you have e.g. two speakers with UK-code tels but one
speaker reveals he is based in another country, this may be a case of PYON and you
could enter 'techref PYON' in B3M. [REDACTED] and [REDACTED] agreed to the
suggestion.
4. B14 / MCS update (courtesy of [REDACTED])
a) Voice
Diariser: over the summer we have successfully deployed version 5.0 of the Diariser.
The main difference between this and version 4.9 is the complete re-write of the
modem arbiter, which now ensures more effective identification of various V series
modems to assist OPC~CAP in de-modulating them.
Speaker ID: since the last VFUG in July we have deployed:
- Two systems for OPI~MENA: one against an Iraqi political target and one against a
Saudi leadership target.
- One system for OPI~RCIT targeting a Gazprom official.
- A 5 way multi-speaker system for OPI~LANG/OPI~SC against Afghan counter
narcotics targets
We have now cleared our order book for speaker id systems and so look forward to a
well deserved, quiet, peaceful and reflective Autumn. Seriously though, if you believe
we can help you with identifying your target of interest amongst the deluge of traffic
that you have to wade through, feel free to approach us and we will happily discuss
your requirements and hopefully offer a swift and accurate solution.
Language ID:
No new deployments. I have however, been concentrating on improving currently
deployed models using non-sigint language data accrued from our Speech to Text
(STT) corpora. Experimentation has shown that it is possible to improve accuracy to
a certain extent by "brute-forcing" the language id algorithm with vast volumes of
data, although this is by no means a substitute for actual "linguist-truthed" sigint
language data accrued from the target set that the system will eventually run against.
It is also quite problematic to service a requirement where we are seeking to identify
one target language from a number of others, some of which are unidentified.
Inevitably this will lead to false positive identifications of the target language as the
algorithm will assume that because a language is not included in the background
3 of 7
This information is exempt from disclosure under the Freedom of Information Act 2000 and may be subject to exemption under
other UK information legislation. Refer disclosure requests to GCHQ on 01242 221491 x30306 (non-sec) or email infoleg@gchq
TOP SECRET STRAP1