Expanding capabilities for AI-powered voice analytics

Lawful intelligence analysts face expanding caseloads, even as the universe of data in each investigation becomes larger and more complex. Efficiencies provided by AI-enabled automation and insight are emerging to close the resulting resource gap, including substantial progress with voice analytics. The environment for lawful interception and processing of audio has been a significant area of change, as traditional mobile voice traffic has become encrypted and over-the-top applications have proliferated. Older methods based on batch processing and rules-based analysis are now being replaced by advanced, flexible real-time services made possible with AI.

The SS8 Intellego XT platform draws on massive open AI models for natural language processing that have each been trained on several hundred thousand hours of audio. In addition to the current capabilities made possible by these models, their open nature ensures that they will continue to benefit from broad-based community development. For example, evolution to include additional languages, dialects, and accents continually makes natural language processing and inference regarding context more robust. Using open models is also critical for operation without internet connectivity, a requirement for deployment in air-gapped environments.

Intellego XT Functionality Overview

The current generation of SS8 voice analytics tools supports semantic understanding of conversations in multiple languages, cross-referenced and correlated with massive universes of investigative data. An increasing breadth of capabilities elevates analyst reach into the potential of voice analytics.

Elevating voice audio as searchable intelligence

Recent, ongoing advances in speech-to-text have dramatically improved accuracy, with the increased fidelity making voice content more consumable, searchable, and faster to triage. Instead of listening to hours of raw audio, analysts can scan transcripts, auto-summarize content, run keyword searches, and quickly identify segments that require deeper review. After pre-processing to scrub noise from the recording, content is automatically rendered into text in real time, with translation from and into multiple languages as needed to overcome language barriers for investigators.

Capabilities built on top of speech-to-text include Intellego XT’s Keyword Spotter, which provides a semantic layer of insight by intelligently monitoring incoming audio to flag specific patterns of keywords or phrases. The underlying AI mechanisms increase the sophistication of content interpretation, such as to distinguish between casual or figurative mentions of words like “bomb” or “kill” versus uses that are criminally meaningful.

The system automatically enforces governance measures such as logging, audit trail, and retention management, with flexible storage policies that allow “high-marked” transcriptions to be designated for extended storage. Administrators can also opt to configure full-text indexing of transcripts to improve accessibility and usability of transcribed data, balancing capability requirements with the use of additional database and storage resources.

The deep integration of speech-to-text with the rest of the SS8 lawful intelligence platform elevates the value of voice content within the investigation as a whole. Real-time monitoring lets analysts set alerts for content of interest without delay and correlate audio with other sources such as location data, device identifiers, communication patterns, and historical case information, both manually and with automated, AI-enabled tools. This fusion is essential to drive actionable insights from complex, intersecting data.

Parsing conversational structure with diarization

Human listeners have the innate ability to parse audio streams and discern which portions belong to individual speakers, which has historically been a shortcoming that limits effectiveness of natural language processing and speech-to-text systems. The accuracy of downstream interpretation depends heavily on the ability to separate and attribute overlapping or sequential voices from raw audio.

The SS8 platform implements sophisticated diarization on monolithic audio streams that breaks the waveform into smaller, speaker-labeled segments for analysis. That capability reduces the potential for analyst confusion and allows for real-time interpretation instead of post-processing, including automated monitoring and insight from voice content.

Segmenting streaming audio by speaker illuminates the interactions among parties in a conversation, including recognizing who says what, tying emotional cues to the correct speaker, and illuminating engagement patterns such as turn-taking and dominance. Diarization is a prerequisite for differentiating individual speakers across languages and dialects, as well as accurate audio analysis in multilingual environments, helping create timelines that map the progress of calls. It also functions as a data normalization layer of sorts, by setting the stage for audio cleanup capabilities such as filtering out silence, removing ambient background noise, and resolving instances where speakers overlap.

Voice recognition can match diarized voice segments to biometric voice profiles, attributing conversation to specific participants in investigations and discerning specific individual targets across multiple incoming audio sources as they occur. Identifying speakers and associating them with their roles in conversations feeds into broader investigative workflows, such as social graph mapping and link charting to visualize patterns of life and relationship dynamics within criminal organizations. The added context of speaker attribution in transcripts provides structure and searchability, sheds light on participants’ intentions, and strengthens the reliability of evidence.

To find out more about SS8’s leading lawful and location intelligence solutions, visit www.SS8.com