Reimea Meridian / Listening agent
Human conversations. Reviewable understanding.
Help the person handling a conversation see what needs attention: the topic, expressed concerns and questions still to address, with the words behind the signals kept visible.
Listen is Meridian’s product for understanding human-to-human conversations. It supports a handler while people talk, helping them notice a repeated concern, a service topic or a question that may still need an answer. The person remains responsible for understanding the context and deciding what to do next.
Use cases include customer-service calls, operational support and authorised healthcare demonstrations. A service owner configures the relevant topics and phrases rather than assuming one vocabulary fits every organisation.
Voice handles agent-led voice interaction and structured capture. Listen analyses the content of conversations between people. Guardian provides supervision and service oversight. They can work together, but each has a distinct purpose.
The reviewed source exposes Listen as a standalone and integrated product service. A production live-audio path still depends on the configured channel, transcription, identity and deployment controls.
Configured phrase matching can surface expressed emotion categories with evidence from the caller’s wording. A label is a prompt to inspect the conversation; it is not a reliable measurement of somebody’s internal emotional state.
Presets cover healthcare, customer service, financial services, utilities, public services and education. Custom topics and phrases allow the service vocabulary to be reviewed before use.
The analysis extracts concern-like sentences and candidate open questions. It also checks repetition and a simple change in wording over time, helping the handler choose what to revisit.
Where enabled, a model can add permitted topic or emotion tags using evidence quotes. The rules-based path remains available, and the optional tags do not change the calculated level or safety routing.
Agree the conversation purpose, participants, permissions and recording or transcription arrangements.
Receive speaker-labelled text through an agreed channel. Inspect transcription quality and speaker attribution.
Apply the service listening profile to the selected utterances, with optional bounded model tags when configured.
Show topics, evidence, concerns and candidate unanswered questions to the handler. Check them against the conversation.
The responsible person clarifies the issue and decides the next step. Any save, handover or escalation needs its own workflow controls.
The implementation combines configurable, explainable rules with optional model-assisted tags. The following describes the supplied source baseline, with deployment boundaries beside each mechanism.
| Layer | Mechanism | Boundary |
|---|---|---|
| Service boundary | The Python ListenService delegates semantic analysis to the shared runtime. Product routes accept utterances or text and obtain the service’s listening profile. | It runs inside the composed application. A separate product name does not imply an independently isolated microservice. |
| Speaker-aware input | The analysis distinguishes caller and handler text. Caller wording drives the surfaced topics and expression signals; handler wording helps the open-question heuristic. | Incorrect speaker labels or transcription errors can alter results. Live channel integration requires its own end-to-end validation. |
| Explainable rules | Phrase lists, limited negation handling, sentence patterns and word counts produce evidence, topics, concerns, repetition and simple wording scores. | These are language heuristics. Sarcasm, accent, context and indirect language can produce missed or misleading signals. |
| Candidate open questions | Question patterns identify candidates. A question can be treated as addressed when the handler uses one of its main words. | Keyword overlap does not establish that an answer was correct, complete or understood. The handler must check. |
| Optional model tags | The voice.semantics task uses a structured schema and configured category IDs. It considers the last four caller utterances, capped at 2,000 characters, and checks evidence against supplied wording. | It adds tags without changing the calculated level or safety routing. Review provider, payload and task settings; metadata alone is not proof that a model completed successfully. |
| Speech and voice signals | The service reports whether Azure Speech configuration is ready and provides helpers for voice-level and combined signals. | Configuration readiness is not a tested live audio connection. Voice or wording scores are not validated psychological, diagnostic or urgency assessments. |
| Outputs and events | Analysis returns semantic fields and speech readiness. The platform transcript route also records a transcript-received conversation event. | Event recording does not by itself establish complete audit coverage, immutable retention or authorisation for every deployment. |
Outputs can include topic labels, expression categories with evidence, concern sentences, candidate open questions, repetition, wording scores, a simple trend and a handler tip. These are advisory signals. Listen does not independently determine clinical urgency, diagnose a condition or establish how somebody feels.
Profiles allow monitoring categories to be switched on or off. Semantic analysis also has a global off switch. Before a pilot, review phrase lists, representative language, speaker attribution, missed signals and false positives with the people who will use the output.
Agree transcript access, retention, participant notices and permitted model tasks. The website is an explanation of the product; it does not capture or analyse visitors’ microphone audio. A live pilot needs an authorised deployment and a measured evaluation of the complete audio-to-review workflow.
Its role is to support understanding of conversations between people. Meridian Voice is the agent-led voice product.
The reviewed semantic analysis has a rules-based path. Optional model assistance adds constrained tags when configured; it is not required for the underlying phrase and question heuristics.
No. The current heuristic can mistake shared words for an answer or miss indirect questions. It helps a handler review the conversation; it does not certify resolution.
No validated clinical or psychological assessment is established by this source review. Signals require context and human review, with intended-purpose and release evidence for any clinical use.
Reviewed 18 October 2026 against the supplied v23.1 source baseline: products/listen/service.py, meridian_runtime/semantics.py, api/product_routes.py and api/platform_routes.py. Source inspection does not establish production readiness.
Read the assurance summary