AI Voice Platforms: Universal Voice Architecture & Engine Registry

Universal Voice Architecture

BrokerOS decouples telecommunications transport from conversational artificial intelligence. Every outbound voice campaign executes across two distinct, independently configured layers:

PSTN Carrier Bridge Layer: Physical telecommunications infrastructure (Vobiz, Exotel, Twilio, Telnyx) that provisions phone numbers, dials target handsets, and manages SIP trunks.
AI Voice Agent Engine Layer: Conversational intelligence platforms (Vapi, Retell, Sarvam, Bolna, ElevenLabs, LiveKit, OpenAI Realtime, Pipecat) that process audio, recognize speech, manage dialog state, and synthesize responses.

This separation prevents vendor lock-in: real estate brokerages can route calls through local Indian PSTN carriers (e.g. Exotel or Vobiz for regulatory TRAI DND compliance) while streaming audio into specialized Indic models (Sarvam) or global ultra-low latency engines (Retell).

Universal Engine Registry

BrokerOS registers all supported voice platforms in integrations/voice/index.ts under the central VOICE_AGENT_REGISTRY:

Platform	Display Name	Category	Native PSTN	Carrier Bridge	Indic Voices	Avg Latency	Target Real Estate Use Cases
VAPI	Vapi AI	Orchestrated Hub	Yes	Yes	Yes	~400 ms	General real estate qualification, CRM variable injection, custom function calling
RETELL	Retell AI	Orchestrated Hub	Yes	Yes	Yes	~350 ms	Ultra-low latency cold calling, high-velocity objection handling, rapid turn-taking
SARVAM	Sarvam AI (India)	Streaming Engine	No	Yes	Yes	~250 ms	Regional Indian language campaigns in Hindi, Tamil, Telugu, Marathi, and Kannada
BOLNA	Bolna AI	Orchestrated Hub	Yes	Yes	Yes	~450 ms	Hinglish code-switching, localized Indian real estate scripts, native Exotel integration
ELEVENLABS	ElevenLabs ConvAI	Streaming Engine	No	Yes	Yes	~300 ms	Ultra-luxury developments (5 Cr+ INR), founder/architect voice cloning, emotional warmth
LIVEKIT	LiveKit Agents	Streaming Engine	No	Yes	Yes	~150 ms	Enterprise WebRTC/SIP room infrastructure, sub-100ms streaming, custom Python agents
OPENAI_REALTIME	OpenAI Realtime API	Streaming Engine	No	Yes	No	~200 ms	Direct speech-to-speech GPT-4o, native inflection detection, natural human interruptions
PIPECAT	Pipecat AI	Streaming Engine	No	Yes	Yes	~180 ms	Open-source modular pipeline, Daily WebRTC, hybrid model routing, on-prem VPC deployments

The IVoiceAgentProvider Contract

All voice agent adapters implement the unified IVoiceAgentProvider interface from packages/types:

export interface IVoiceAgentProvider {
  readonly platformType: VoiceAgentPlatform;
  
  validateCredentials(credentials?: VoiceAgentCredentials): Promise<boolean>;
  
  previewAudio(
    text: string,
    voiceId: string,
    credentials?: VoiceAgentCredentials
  ): Promise<{ audioBuffer: Buffer; contentType: string }>;
  
  dispatchOutboundCall(
    options: SendVoiceOptions,
    credentials?: VoiceAgentCredentials
  ): Promise<SendVoiceResult>;
  
  parseWebhookEvent(headers: Record<string, any>, payload: any): VoiceWebhookEvent[];
  
  getAvailableModels(credentials?: VoiceAgentCredentials): Promise<VoiceModelItem[]>;
  
  getAvailableVoices(credentials?: VoiceAgentCredentials): Promise<VoicePersonaItem[]>;
}

Carrier Bridge Dispatcher Logic

When a voice campaign starts, tryCarrierBridgeDispatch evaluates whether PSTN carrier credentials (Vobiz, Exotel, Twilio, or Telnyx) are present. If configured, the call is placed through the carrier’s PSTN trunk, and an answer_url webhook is passed to connect the answered call directly to the chosen AI agent:

export async function tryCarrierBridgeDispatch(
  platform: VoiceAgentPlatform,
  options: SendVoiceOptions,
): Promise<CarrierBridgeResult> {
  const telCreds = options.telephonyCredentials;
  if (!telCreds) {
    return { handled: false };
  }

  const prov = (telCreds.provider || '').toUpperCase();
  // Evaluates VOBIZ, EXOTEL, TWILIO, or TELNYX and routes to PSTN trunk
  // with dynamic answer_url pointing back to the BrokerOS voice gateway.
}

Monorepo Workspace Package Map

Voice integrations live within self-contained packages in the monorepo root:

Aggregator Module: integrations/voice/index.ts
Vapi: integrations/voice/agents/vapi/ (@brokeros/int-agent-vapi)
Retell AI: integrations/voice/agents/retell/ (@brokeros/int-agent-retell)
Sarvam AI: integrations/voice/agents/sarvam/ (@brokeros/int-agent-sarvam)
Bolna: integrations/voice/agents/bolna/ (@brokeros/int-agent-bolna)
ElevenLabs: integrations/voice/agents/elevenlabs/ (@brokeros/int-agent-elevenlabs)
LiveKit: integrations/voice/agents/livekit/ (@brokeros/int-agent-livekit)
OpenAI Realtime: integrations/voice/agents/openai-realtime/ (@brokeros/int-agent-openai-realtime)
Pipecat: integrations/voice/agents/pipecat/ (@brokeros/int-agent-pipecat)

Setup Steps

Configure Telephony Carrier

Navigate to Settings -> Integrations -> Telephony Carriers and connect your PSTN trunk (e.g. Vobiz, Exotel, Twilio, or Telnyx).

Connect AI Voice Platform

Navigate to Settings -> Integrations -> AI Voice Agents -> Add Platform and supply the API key for your chosen voice engine (e.g. Vapi, Retell, Sarvam).

Compose Voice Campaign

In the Voice Campaign Wizard, choose your target project, select the telephony carrier for dialing, and configure the voice persona, script, and variables in the Agent Composer Studio.

Getting Started

Phone Carriers (PSTN)