Neurodivergent-friendly voice UX (not moderation)

Calm the room without changing the conversation

CalmCue reduces sensory overload in voice chat by adapting audio dynamics in real time: leveling loudness, reducing overlap, and offering gentle turn-taking cues.
No toxicity detection. No content classification. Only overlap, interruptions, and loudness spikes.

Overlap-aware

Detects sustained overlap and helps you hear the dominant speaker.

Volume-stable

Smooth leveling reduces sudden loudness spikes and jumps.

Private recaps

Opt-in summaries help you rejoin after overload moments.

The problem

Voice rooms get chaotic fast

In many voice chats, the hard part isn’t what people are saying—it’s the audio dynamics: overlap, interruptions, and sudden loudness changes. For a lot of people (especially neurodivergent folks), that turns into fatigue, anxiety, or “I can’t follow this anymore.”

Overload comes from dynamics

Not content

When multiple people talk at once, the brain does extra work just to parse who’s speaking and what to track.

  • Sustained overlap makes speech unintelligible
  • Interruptions break turn-taking expectations
  • Loudness spikes trigger stress/avoidance

Most tools focus on moderation

We don’t

CalmCue is not a toxicity detector. It doesn’t interpret meaning. It just smooths the room so you can stay present.

  • No sentiment analysis
  • No topic or intent classification
  • Just signals → comfort outputs
Goal

Keep the conversation intact, but make it easier on the nervous system.

Core features

Calmer rooms, same conversation

Reduce cognitive load from chaotic voice chat—without changing what anyone is allowed to say.

Chaos meter

Audio dynamics

Tracks overlap ratio, interruptions, and loudness spikes in real time. Converts signals into an explainable 0–100 score.

  • Overlap ratio (who’s talking over whom)
  • Interruptions count (turn-taking breaks)
  • Loudness spikes (sudden volume jumps)

Overlap nudge

Gentle & rate-limited

When overlap persists, CalmCue offers a gentle toast: “Let Speaker A finish before Speaker B.”

  • Triggered only by overlap duration
  • Cooldown to avoid nagging
  • No analysis of words or topics

Dynamic leveling + ducking

Comfort

Smoothly levels loudness and ducks the non-dominant speaker during sustained overlap to improve intelligibility.

  • Volume leveling to a comfort target
  • Ducking during sustained overlap
  • Configurable strength + thresholds

Focus mode recap

Private

If chaos stays high, CalmCue can send you a private recap of the last 15–120 seconds—so you can rejoin without stress.

  • Private DM-style card
  • Time-window selectable
  • Optional LLM recap (or heuristic fallback)
How it works

Signals in, comfort out

A lightweight pipeline that stays away from content and focuses on dynamics.

01

Measure dynamics

RMS → dB, voice activity, rolling baseline, overlap windows.

02

Compute chaos

Weighted sum of overlap, interruptions, spikes → 0–100 score.

03

Apply shields

Leveling, ducking, nudges, and optional recap prompts.

Explainable scoring

chaosScore = 55% overlap + 25% interruptions + 20% loudness spikes (scaled & capped). No “meaning” is inferred from speech.

FAQ

Common questions

Clear boundaries, simple answers.

Is this moderation or toxicity detection?

No. CalmCue does not analyze content, intent, sentiment, or policy. It only measures overlap/interruptions/loudness dynamics.

Does CalmCue record or store audio?

This site positions CalmCue as dynamics-first. Your product can process audio locally when possible and store only aggregated telemetry.

Is the recap visible to everyone?

No. The recap is private to the user who requests it (DM-style), not posted to the room.

Who is it for?

Anyone who gets overwhelmed by chaotic voice rooms—especially neurodivergent folks, anxious listeners, and accessibility-first teams.

Self-learning policy

Feedback that updates the room—safely

CalmCue can adapt its dynamics settings based on lightweight feedback—without analyzing what was said. Updates are rate-limited, clamped, and applied only at session end.

In-session feedback

“Too aggressive” / “Too weak”

We aggregate feedback during the session, then apply a single policy update only when the session ends (/api/session/end).

  • Feedback is collected during the call
  • Policy updates apply only at session end
  • UI shows policy vN + a short explanation of what changed

Safety rails

Clamped + max 10%

Parameter changes are clamped to safe ranges, with a maximum of 10% change per session to prevent oscillation or “surprise” behavior.

  • Clamp parameters to allowed min/max
  • Max 10% delta per session
  • Change summary shown to the user

Example updates

Dynamics only

What feedback can adjust (examples):

  • Too aggressive → ↑ k, ↑ tSec, ↓ duckingStrength, ↑ toastCooldownMs
  • Too weak → ↓ k, ↓ tSec, ↑ duckingStrength, ↓ toastCooldownMs

Persistence

Postgres + local cache

Each update creates a new Policy vN in Postgres, and caches the latest policy in localStorage so the next run starts from your last known-good settings.

  • Create Policy vN in Postgres
  • Cache latest in localStorage
  • Restore on next launch
Boundary

Transcripts are not used for policing. Any text is used only for recap summarization when the user opts in.

Optional integrations

Tools we plugged in (and why)

CalmCue stays dynamics-first. These integrations add captions, private recaps, and real observability—without turning into content policing.

Modulate — Velma Transcribe (captions)

Captions

Modulate’s Velma Transcribe gives CalmCue live captions and a rolling transcript with speaker labels and timestamps. We use that transcript only to generate the user’s private recap when they opt in—never for moderation, filtering, or policing.

Airia Gateway — Focus mode recap

Private recap

When chaos stays high for a bit, CalmCue can offer a private recap to help you rejoin. We send only the last 15/30/60/120 seconds (your choice) for summarization, and fall back to a lightweight heuristic if needed.

Braintrust — tracing, logs, eval-ready

Observability + improvement

We log focus summaries and session-end events as traces so you can debug any “why did it do that?” moment and build an eval dataset over time.

  • Inputs + outputs + scores per run
  • Find any session fast (sessionId)
  • Same logs become your eval dataset later

Lightdash — product dashboards

Metrics & trends

We track how CalmCue behaves over time: reward trends, overlap seconds per policy version, and “first vs second run” improvements after policy learning.

  • Reward trend
  • Average overlap seconds per policy version
  • First vs second run comparisons
Privacy by design

Integrations are opt-in. CalmCue’s core behavior never depends on content understanding—only dynamics signals.

Future scope

From a demo to real teams and real rooms

CalmCue’s core idea works anywhere people talk over each other—so the natural next step is making it easy to plug into the tools teams already live in.

Team rollout

Workspaces

Shared defaults with personal comfort profiles—so individuals can opt into what helps them without forcing it on everyone.

  • Workspace-level policy templates
  • Per-user “comfort mode” preferences
  • Admin controls + privacy boundaries

Zoom / Meet / Teams / rooms

Integrations

Bring dynamics smoothing, gentle turn-taking cues, and private recaps into existing voice/video calls.

  • Client-side or companion-app audio processing
  • Meeting recap cards per user (opt-in)
  • Room-level analytics (aggregate only)

Richer “comfort controls”

Personalization

More control over what “calm” means for you—without turning it into moderation.

  • Adjust sensitivity for overlap vs loudness
  • Different cue styles (toast, subtle UI, haptics)
  • Context modes (standup, brainstorming, social)

Evaluation + improvement loops

Evals

Use Braintrust logs + user feedback to test changes safely and prove “this actually reduced overload.”

  • Offline eval sets for summarization
  • A/B tests for nudges and thresholds
  • Policy versioning + measurable outcomes
North star

CalmCue becomes a “comfort layer” for voice/video—integrated into the places people already meet, without changing the conversation itself.