Skip to main content

Voice Nox

Audience: Sales · Support · QA · Engineering · Management · Where in app: Dashboard → Nox → Voice · Plan availability: Pro+ (a further entitlement on top of base Nox — verify)

Voice Nox lets an admin talk to Nox instead of typing. It's the same assistant with the same tools, but the conversation is spoken — useful while driving, on mobile, or for a quick check while doing something else. It's powered by ElevenLabs Conversational AI over a WebSocket.

What it does​

  • Answers the same questions as text Nox, spoken aloud.
  • Can trigger the same tools (visitor lookups, stats, knowledge search, place a call, etc.).
  • Captures a transcript of the voice session for audit.

Voice-friendliness varies by answer shape:

QuestionVoice-friendly?
"How many bookings yesterday?"Yes — single number
"Walk me through John's journey"Yes — natural narrative
"Show me Sarah's last 5 visits"Partial — summarized verbally, full detail in chat
"Which 17 leads should I prioritize?"Partial — top few verbal, full list in chat

Long tables / multi-row lists are summarized for voice; the admin can ask for "the full list" to get it in the chat panel instead.

How it works​

  1. Admin opens Voice Nox in the dashboard.
  2. Frontend calls POST /knox-voice/signed-url on ms-ai.
  3. ms-ai mints a signed WebSocket URL — a synced ElevenLabs agent with the tenant's company context baked in.
  4. The browser opens the WebSocket; audio streams bidirectionally.
  5. The ElevenLabs agent invokes Nox tools as needed via the WebSocket tool-call protocol (WebSocket → ms-ai → tool registry → result → back to ElevenLabs → narrated).
  6. Audio response streams back. During-call display events (canvas cards, sources, tool results) stream to the browser over GET /knox-voice/events/:conversationId (SSE), fanned out via Redis pub/sub (knox-voice-events.service.ts) — the ElevenLabs tool webhook and the browser subscription land on different pods in production.
  7. On call end, the transcript is captured to knox_call_transcripts.

A voice call continues the shared Nox conversation: the frontend passes the recent chat turns and knox-voice.service.ts injects a continuity line, so Nox picks up where the chat left off instead of re-introducing itself.

Tool-call latency is typically 1–2s.

Agent sync​

The Nox voice agent is a synced copy of a base ElevenLabs Conversational AI agent, with per-tenant overrides:

  • Tool whitelist
  • Tenant-specific system-prompt additions
  • Voice ID

Syncing is handled by KnoxVoiceService (ms-ai).

Configuration & options​

  • Entitlement — Voice is gated separately from text Nox.
  • Voice ID / agent overrides — per tenant via the synced ElevenLabs agent.
  • Tool whitelist — which tools the voice agent may call.
  • Recording — audio recording is optional per provider settings; the transcript is always stored for audit.

Behaviors & edge cases​

  • Place a call: "call Sarah" works if the knox_place_call entitlement is on — voice confirms, then the outbound call is queued (or scheduled via scheduled_at).
  • Book a meeting: on Nox funnel beta tenants, knox_book_meeting books a GHL calendar slot after an explicit confirmation; other tenants still can't.
  • White-label: the agent self-references the tenant's ai_assistant_name and never says "Knock Knock" on white-label deployments; agencies can override the TTS voice via agency.nox_voice_id (applied per session in knox-agent-sync.service.ts).
  • Memory: same per-session memory model as text Nox (fresh each session; durable per-admin facts via knox_remember apply to both).
  • Different LLM: the voice agent uses a low-latency model tuned for conversation, so answers can differ subtly from text Nox.
  • Network sensitivity: ElevenLabs needs a stable WebSocket — flaky networks degrade quality fast, and complex multi-tool queries feel slow over voice.

Plan & limits​

  • Plan-gated (Pro+, separate entitlement — verify).
  • Consumes AI credits per minute of conversation.
  • Some tool results don't translate to voice and are summarized — full detail is available in the chat panel.

Privacy​

  • The voice transcript is stored in knox_call_transcripts for audit.
  • Audio recording is optional per provider settings.
  • The admin can delete a transcript via the standard delete flow.

Technical implementation​

  • Module: ms-ai/src/knox-voice/. Signed URL: POST /knox-voice/signed-url. Agent sync: KnoxVoiceService.
  • Tools are the same registry as text Nox (knox-agent/tools/), bridged over the ElevenLabs WebSocket tool-call protocol.
  • Transcripts land in knox_call_transcripts (ms-ai Postgres) — see Visitor data.