first commit
This commit is contained in:
commit
0d90363116
6510 changed files with 679146 additions and 0 deletions
|
|
@ -0,0 +1,322 @@
|
|||
TOP SECRET
|
||||
|
||||
Innov8 Experiment Profile
|
||||
|
||||
| | | | |
|
||||
| :------------------ | :-------------- | :------------------- | :------------ |
|
||||
| Experiment Name | Voice Analytics | Experiment Reference | |
|
||||
| Version | 1.0 | Date | 29 March 2010 |
|
||||
| Experiment Owner | [REDACTED] | Department | T12 |
|
||||
| Experiment Start Date | 5 July 2010 | Experiment End Date | August 2010 |
|
||||
|
||||
Part 1 - Experiment Overview
|
||||
|
||||
TOP SECRET
|
||||
|
||||
TOP SECRET
|
||||
|
||||
Business Case
|
||||
|
||||
Innov8 Experiment Profile Template V0.2
|
||||
2 of 13
|
||||
|
||||
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 [REDACTED] or email [REDACTED]
|
||||
|
||||
TOP SECRET
|
||||
|
||||
TOP SECRET
|
||||
|
||||
Detail the nature of the Experiment as a high level plain English description. Refrain from using
|
||||
overly technical terms if possible. Justify briefly why this Experiment should run, with a high level
|
||||
summary of Experiment, Conditions, De-risking effort, Business Case and Expected Outcomes/
|
||||
Benefits.
|
||||
|
||||
As part of the Joint Capability Activity (JCA) in partnership with GCHQ, experiments
|
||||
by Voice Analytics are to be conducted within the Joint Collaboration Environment
|
||||
(JCE) using DISTILLERY. VoiceRT is NSA's currently deployed voice analytics
|
||||
technology that delivers integrated speaker ID (SID), language ID (LID), gender ID
|
||||
(GID), dual-tone multi-frequency (DTMF) detection, speech activity detection
|
||||
(SAD), and phonetic keyword search (i.e., direct keyword search on audio data) in
|
||||
more than 25 key foreign languages. VoiceRT requires specific hardware,
|
||||
proprietary process provisioning software (Tibco), and uses an older voice analytics
|
||||
technology. A new initiative is underway to create the next generation of VoiceRT
|
||||
that operates on commodity hardware (GHOSTMACHINE), replaces the Tibco
|
||||
process provisioning software with DISTILLERY, and makes use of the latest voice
|
||||
analytics technology from R64 (including the new Speech-to-Text). Voice Analytics
|
||||
experiments will occur in phases with this proposal being the first phase.
|
||||
|
||||
The business case for running Phase 1 of the Voice Analytics experiment in the
|
||||
JCE is to demonstrate the following:
|
||||
* LPT voice data can be ingested into GHOSTMACHINE and passed to
|
||||
DISTILLERY for processing at the rate of arrival.
|
||||
* DISTILLERY can run a subset of voice analytics (SAD, GID, and LID) on
|
||||
voice data.
|
||||
* Running multiple analytics on voice data in parallel produces the same result
|
||||
as running each analytic independently.
|
||||
* DISTILLERY can pass results from an analytic as input to a follow-on
|
||||
analytic with decision points throughout the dataflow based on analytic
|
||||
results (e.g. If GID reports the speaker is male, pass the data to LID;
|
||||
otherwise, stop processing the voice cut).
|
||||
* DISTILLERY is capable of resource management at high speeds. If a voice
|
||||
analytic process in DISTILLERY is unable to handle the ingest rate,
|
||||
DISTILLERY can spawn additional voice analytic processes. Similarly, if
|
||||
resources from one set of voice analytic processes can be better used
|
||||
elsewhere, DISTILLERY will remove from memory some/all of the set of
|
||||
voice analytic processes.
|
||||
* Determine the scalability as additional analytics are loaded into memory and
|
||||
processing data. How is the previous bullet affected as free memory goes to
|
||||
zero? How quickly can voice analytic processes be spawned and broken
|
||||
down?
|
||||
|
||||
Innov8 Experiment Profile Template V0.2
|
||||
|
||||
3 of 13
|
||||
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 [REDACTED] or email [REDACTED]
|
||||
|
||||
TOP SECRET
|
||||
|
||||
TOP SECRET
|
||||
|
||||
Benefits
|
||||
|
||||
| Outcome | Expected Benefit | How will benefit be measured? | KPI (CSF) | Reference Number |
|
||||
| :--- | :--- | :--- | :--- | :--- |
|
||||
| Demonstrate the JCE environment (including GHOSTMACHINE and DISTILLERY) can get data to core analytics at ingest rate. | The JCE environment can be used to process voice content. GHOSTMACHINE and DISTILLERY can maintain speeds required to processes voice on a 10G link. Add DISTILLERY to GHOSTMACHINE stack. | Monitor ingest pick-up directory that feeds DISTILLERY input. | | |
|
||||
| Demonstrate the JCE architecture can affordably scale in line with other VoIP accesses. | If the outcome is true, then the JCE environment provides a good test environment for future voice processing experiments. If VoiceRT is to eventually be replaced with this new technology, then tests must be run against a similar environment. This experiment also tests DISTILLERY'S ability to dynamically manage resources, which is a requirement for our voice processing. | Compare quantity of VoIP data processed at JCE against that collected on similar links at NSAW and GCHQ. | | |
|
||||
| Demonstrate voice analytic processing of incoming data significantly reduces the amount of unintended collection reaching long-term stores by filtering on Language ID, Speech Activity Detection, and Gender ID. | Analysts can be directed to smaller, more probable collected voice cuts. Since voice files can be very large, significantly reduces the amount of long-term storage required. | DISTILLERY will be configured to separate wanted and unwanted VoIP files in separate storage locations. Comparing sizes of the two locations will provide an approximation for the ratio of wanted versus unwanted traffic. | | |
|
||||
|
||||
Innov8 Experiment Profile Template V0.2
|
||||
|
||||
4 of 13
|
||||
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 [REDACTED] (non-sec) or email [REDACTED]
|
||||
|
||||
TOP SECRET
|
||||
|
||||
TOP SECRET
|
||||
|
||||
| | | | | |
|
||||
| :--- | :--- | :--- | :--- | :--- |
|
||||
| Demonstrate voice processing can be performed on commodity hardware (GHOSTMACHINE). Current voice processing at NSA has hardware specific requirements. | Voice analytics can be executed on commodity hardware alongside other applications. Allows for easier hardware maintenance and shorter deployments. | Analysts will confirm resulting data has been filtered without a loss of wanted traffic. | | |
|
||||
| JCE experiments can process content while remaining in compliance with US and UK laws. | Future experiments requiring content for processing can leverage. | If MONKEYPUZZLE can safely produced selected content to the experiment, then it should be in compliance. | | |
|
||||
| | | | | |
|
||||
| | | | | |
|
||||
|
||||
Innov8 Experiment Profile Template V0.2
|
||||
|
||||
5 of 13
|
||||
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 [REDACTED] (non-sec) or email [REDACTED]
|
||||
|
||||
TOP SECRET
|
||||
|
||||
TOP SECRET
|
||||
|
||||
Plan
|
||||
|
||||
Timeline, including dependencies/pre-requisites, set-up, operation, evaluation, reporting,
|
||||
closedown. Also includes deliberate outages of equipment.
|
||||
|
||||
As stated in the business case, this is Phase 1 of the Voice Analytics Experiment.
|
||||
This phase is divided into four sub-experiments focusing on dataflow through a select
|
||||
group of voice analytics (SAD, GID, LID) in DISTILLERY. All four experiments
|
||||
require access to CSDF voice files. Future phases will consider SOTF once
|
||||
GHOSTMACHINE can support this format.
|
||||
|
||||
Exp. 1: Independent Analytic Tests (~1 week) - Designed to test the dataflow from
|
||||
DISTILLERY input, through a few processing blocks, and the output displayed to a
|
||||
GUI. Each analytic is executed independently as a single instance to determine
|
||||
accuracy and processing speed.
|
||||
|
||||
```
|
||||
+------------------------------------+
|
||||
| Distillery |
|
||||
| |
|
||||
CSDF | +------------+ |
|
||||
Voice====|===========>| CSDF | |
|
||||
File | | Ingester | |
|
||||
| +-----^------+ |
|
||||
| | |
|
||||
| v |
|
||||
| +--------------+ |
|
||||
| | Configuration| |
|
||||
| | Retrieval | |
|
||||
| +-----^--------+ |
|
||||
| | |
|
||||
| v |
|
||||
| +------------+ |
|
||||
| | Analytic* | =========|==> GUI
|
||||
| +------------+ |
|
||||
+------------------------------------+
|
||||
```
|
||||
|
||||
* Speech Activity Detection (SAD), Language ID (LID), and Gender ID (GID)
|
||||
will each be tested independently as the active analytic
|
||||
|
||||
Innov8 Experiment Profile Template V0.2
|
||||
|
||||
6 of 13
|
||||
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 [REDACTED] (non-sec) or emai[REDACTED] chq
|
||||
|
||||
TOP SECRET
|
||||
|
||||
TOP SECRET
|
||||
|
||||
Exp. 2: Parallel Analytic Tests (~1 week) - Similar to first experiment, but each
|
||||
CSDF voice file is run against each of the three voice analytics simultaneously. This
|
||||
should confirm data is not corrupted by multiple processes accessing simultaneously.
|
||||
Experiment also adds the complexity of holding results until all analytics have
|
||||
executed prior to displaying on GUI.
|
||||
|
||||
```
|
||||
+--------------------------------------------------------+
|
||||
| Distillery |
|
||||
| |
|
||||
CSDF | +------------+ |
|
||||
Voice====|===========>| CSDF | |
|
||||
File | | Ingester | |
|
||||
| +-----^------+ |
|
||||
| | |
|
||||
| v |
|
||||
| +--------------+ |
|
||||
| | Configuration| |
|
||||
| | Retrieval | |
|
||||
| +-----^--------+ |
|
||||
| | |
|
||||
| +----------------------+----------------+ |
|
||||
| v v v |
|
||||
| +----------------+ +----------------+ +----------------+
|
||||
| | Speech Activity| | Language ID | | Gender ID |
|
||||
| | Detection (SAD)| | (LID) | | (GID) |
|
||||
| +----------------+ +----------------+ +----------------+
|
||||
|
|
||||
|
|
||||
v
|
||||
GUI
|
||||
```
|
||||
|
||||
Exp. 3: Linked Analytic Tests (~1 week) - Similar to the previous experiments, but
|
||||
results from the voice analytics are now used to make decisions about future
|
||||
processing of each file. The ingest rate should be fast enough to create a noticeable
|
||||
bottleneck in flow.
|
||||
|
||||
```
|
||||
CSDF
|
||||
Voice
|
||||
File
|
||||
|
|
||||
+---------------------------------------------------------+
|
||||
| | Distillery |
|
||||
| | |
|
||||
| v +---------+ |
|
||||
| +-------------+ +-------------+ |Speech | |
|
||||
| |CSDF Ingester+--->|Configuration| |Activity | |
|
||||
| +-------------+ |Retrieval +------->|Detection| |
|
||||
| +-------------+ |(SAD) | |
|
||||
| +----+----+ |
|
||||
| | |
|
||||
| +----------------+ | |
|
||||
| |Speech Activity | | |
|
||||
| |Above Threshold?|<-----------+ |
|
||||
| +-------+--------+ |
|
||||
| v YES |
|
||||
| +---------------+ |
|
||||
| |Gender ID (GID)| |
|
||||
| +-------+-------+ |
|
||||
| v |
|
||||
| +-------------+ |
|
||||
| |Male Speaker?| |
|
||||
| +-------+-----+ |
|
||||
| v YES |
|
||||
| +-----------------+ |
|
||||
| |Language ID (LID)| |
|
||||
| +-------+---------+ |
|
||||
| v |
|
||||
| +-----------------+ |
|
||||
| |Is It Language X?| |
|
||||
| +-------+---------+ |
|
||||
| | |
|
||||
| | |
|
||||
+---------------------------+-----------------------------+
|
||||
| YES
|
||||
v
|
||||
+-----+
|
||||
| GUI |
|
||||
+-----+
|
||||
```
|
||||
|
||||
Innov8 Experiment Profile Template V0.2
|
||||
|
||||
7 of 13
|
||||
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 o[REDACTED] (non-sec) or emai[REDACTED] chq
|
||||
|
||||
TOP SECRET
|
||||
|
||||
TOP SECRET
|
||||
|
||||
Exp. 4: Dynamic Resource Manager Tests (~2 week) - Builds off of the third
|
||||
experiment. The SAD algorithm is much faster than GID and LID. In this
|
||||
experiment, DISTILLERY should be able to dynamically create/remove additional
|
||||
GID and LID analytic processes to compensate for GID and LID processes falling
|
||||
behind or becoming idle. A more advanced part of this experiment would be to
|
||||
monitor DISTILLERY's response to all system memory in use; for example, 100 GID
|
||||
analytics and 150 GID analytics are still falling behind, but no system memory exists
|
||||
to spawn additional processes. Does one set of analytic processes starve or do we
|
||||
see a constant exchange of memory between GID and LID?
|
||||
|
||||
```
|
||||
CSDF
|
||||
Voice
|
||||
File
|
||||
|
|
||||
+---------------------------------------------------------+
|
||||
| | Distillery |
|
||||
| | |
|
||||
| v +---------+ |
|
||||
| +-------------+ +-------------+ |Speech | |
|
||||
| |CSDF Ingester+--->|Configuration| |Activity | |
|
||||
| +-------------+ |Retrieval +------->|Detection| |
|
||||
| +-------------+ |(SAD) | |
|
||||
| +----+----+ |
|
||||
| | |
|
||||
| +----------------+ | |
|
||||
| |Speech Activity | | |
|
||||
| |Above Threshold?|<-----------+ |
|
||||
| +-------+--------+ |
|
||||
| v YES |
|
||||
| +---------------+ |
|
||||
| |Gender ID (GID)| |
|
||||
| +-------+-------+ |
|
||||
| v |
|
||||
| +-------------+ |
|
||||
| |Male Speaker?| |
|
||||
| +-------+-----+ |
|
||||
| v YES |
|
||||
| +-----------------+ |
|
||||
| |Language ID (LID)| |
|
||||
| +-------+---------+ |
|
||||
| v |
|
||||
| +-----------------+ |
|
||||
| |Is It Language X?| |
|
||||
| +-------+---------+ |
|
||||
| | |
|
||||
| | |
|
||||
+---------------------------+-----------------------------+
|
||||
| YES
|
||||
v
|
||||
+-----+
|
||||
| GUI |
|
||||
+-----+
|
||||
```
|
||||
|
||||
Innov8 Experiment Profile Template V0.2
|
||||
|
||||
8 of 13
|
||||
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 [REDACTED] (non-sec) or email [REDACTED] gchq
|
||||
|
||||
TOP SECRET
|
||||
|
|
@ -0,0 +1,57 @@
|
|||
{
|
||||
"title": "Innov8 Voice Analytics Experiment Profile",
|
||||
"type": "REPORTED",
|
||||
"description": "This document outlines Phase 1 of a Voice Analytics experiment conducted by GCHQ in partnership with the NSA, testing the ingestion and processing of voice data through the DISTILLERY platform on GHOSTMACHINE hardware. The experiments evaluate independent, parallel, and linked analytics (SAD, GID, LID), as well as dynamic resource management, to ensure accurate, scalable, and compliant voice processing in a Joint Collaboration Environment.",
|
||||
"long_description": "This document details Phase 1 of a Voice Analytics experiment conducted within the Joint Collaboration Environment (JCE) in partnership between GCHQ and the NSA. The experiment is designed to test the ingestion and processing of voice data using the DISTILLERY platform on GHOSTMACHINE hardware, a system intended to replace the older VoiceRT technology. VoiceRT provides integrated speaker ID, language ID, gender ID, DTMF detection, speech activity detection, and phonetic keyword search across more than 25 foreign languages, but it relies on specialized hardware and proprietary software. The new initiative aims to modernize this capability by running advanced voice analytics on commodity hardware, improving scalability, and leveraging updated speech-to-text and analytic algorithms from R64. The first phase of the experiment focuses on demonstrating that voice data can flow through the system at high speeds while maintaining analytic accuracy and reliability.\n\nThe experimental plan is divided into four sub-experiments, each designed to test a different aspect of the voice analytics pipeline. The first experiment evaluates independent analytic performance by running SAD, GID, and LID individually to assess accuracy and processing speed. The second experiment tests the system’s ability to run multiple analytics in parallel without data corruption. The third experiment links analytics so that outputs from one analytic inform decisions for subsequent processing, simulating real-world decision-making workflows. Finally, the fourth experiment examines dynamic resource management, where DISTILLERY spawns or removes analytic processes in response to system load and memory availability, ensuring that processing bottlenecks are avoided and resources are efficiently allocated.\n\nThe expected outcomes of these experiments include demonstrating that the JCE environment can process incoming voice data at ingest rates, scaling affordably in line with other VoIP access requirements, and filtering unwanted content while reducing storage needs. The experiments also aim to validate that voice analytics can run on commodity hardware without loss of data fidelity. Successful completion of Phase 1 will provide the foundation for subsequent phases, which may expand analytic coverage and incorporate additional voice data formats, ultimately improving operational efficiency and enabling more effective real-time analysis of voice communications.",
|
||||
"tags": [
|
||||
"TIBCO",
|
||||
"Innov8",
|
||||
"voice analytics",
|
||||
"VoiceRT",
|
||||
"DTMF",
|
||||
"speech activity detection",
|
||||
"SAD",
|
||||
"phonetic keywords",
|
||||
"Joint Capability Activity",
|
||||
"JCA",
|
||||
"Joint Collaboration Environment",
|
||||
"JCE",
|
||||
"voice processing",
|
||||
"resource scaling",
|
||||
"VoIP"
|
||||
],
|
||||
"document": {
|
||||
"creator": "GCHQ",
|
||||
"date_created": "2010-03-29",
|
||||
"classification": [
|
||||
"TOP SECRET"
|
||||
],
|
||||
"dissemination": [],
|
||||
"banner": "TOP SECRET",
|
||||
"genre": "report",
|
||||
"programs": [
|
||||
"DISTILLERY",
|
||||
"VoiceRT",
|
||||
"GHOSTMACHINE",
|
||||
"MONKEYPUZZLE"
|
||||
],
|
||||
"countries": [],
|
||||
"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_source": "https://theintercept.com/document/2018/01/19/innov8-voiceanalytics-experiment-profile/",
|
||||
"doc_url": "https://s3.documentcloud.org/documents/4352010/Innov8-VoiceAnalytics-Experiment-Profile-v1-0.pdf"
|
||||
}
|
||||
],
|
||||
"pdf_sha256": "c5a3f4c41e7a767c2804286566f26d7b26b3afa531c20d9b4acf6febe728d4e2",
|
||||
"pages": 8
|
||||
}
|
||||
Loading…
Add table
Add a link
Reference in a new issue