Voice by aiConnected — A Developer’s Introduction

What This Document Is

This document explains Voice by aiConnected for developers who are new to voice AI, real-time systems, or this specific project. It covers what we’re building, how the pieces fit together, and what you need to understand to contribute effectively. If you’re joining this project and want to get up to speed quickly, start here.

What We’re Building

The One-Sentence Version

A platform that connects phone calls to AI agents that can have natural conversations, take actions, and transfer to humans when needed.

The One-Paragraph Version

Voice by aiConnected is a multi-tenant Voice AI platform. When someone calls a business using our service, the call is answered by an AI that listens (using speech-to-text), thinks (using a large language model), and responds (using text-to-speech). The AI has access to information about the business and can perform actions like scheduling appointments or updating CRM records. When the AI can’t handle something, it seamlessly transfers the call to a human.

The Visual Version

┌──────────────────────────────────────────────────────────────────────────────┐
│                          THE VOICE PIPELINE                                  │
│                                                                              │
│   Caller                                                                     │
│     │                                                                        │
│     │ Phone Call (audio)                                                     │
│     ▼                                                                        │
│   ┌─────────────┐                                                            │
│   │ GoToConnect │  ← Traditional phone system (PBX)                          │
│   │   (PSTN)    │                                                            │
│   └──────┬──────┘                                                            │
│          │                                                                   │
│          │ WebRTC (audio over internet)                                      │
│          ▼                                                                   │
│   ┌─────────────┐                                                            │
│   │   WebRTC    │  ← Bridges phone audio to our system                       │
│   │   Bridge    │                                                            │
│   └──────┬──────┘                                                            │
│          │                                                                   │
│          │ Audio stream                                                      │
│          ▼                                                                   │
│   ┌─────────────┐                                                            │
│   │   LiveKit   │  ← Real-time communication infrastructure                  │
│   │    Room     │                                                            │
│   └──────┬──────┘                                                            │
│          │                                                                   │
│          ├──────────────────┬──────────────────┬────────────────────┐       │
│          │                  │                  │                    │       │
│          ▼                  ▼                  ▼                    ▼       │
│   ┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐  │
│   │  Deepgram   │    │   Claude    │    │ Chatterbox  │    │  Webhooks   │  │
│   │    (STT)    │───▶│   (LLM)     │───▶│   (TTS)     │    │   (Tools)   │  │
│   └─────────────┘    └─────────────┘    └─────────────┘    └─────────────┘  │
│                                                                              │
│   "Hi, I need to       "The customer        "Your appointment            │
│    schedule an          wants to book        is confirmed for              │
│    appointment"         an appointment.      Tuesday at 3 PM"              │
│                         Let me check                                        │
│                         availability..."                                    │
│                                                                              │
└──────────────────────────────────────────────────────────────────────────────┘

Core Concepts You Need to Understand

1. The Voice Pipeline

Voice AI is essentially a pipeline that transforms audio → text → response → audio:

Audio In → STT → Text → LLM → Response Text → TTS → Audio Out

Each step has latency, and latency is the enemy. If the AI takes too long to respond, the conversation feels unnatural. Our target is under 1 second from when the caller stops speaking to when they hear the AI’s response.

Stage	What Happens	Target Latency
STT (Speech-to-Text)	Converts caller’s voice to text	~300ms
LLM (Large Language Model)	Generates a response	~350ms
TTS (Text-to-Speech)	Converts response to speech	~150ms
Network/Audio	Moving audio around	~100ms
Total		~900ms

2. Streaming vs. Batch Processing

The key to achieving low latency is streaming. Instead of waiting for each stage to complete fully before starting the next, we stream data through the pipeline: Batch (slow):

[Wait for caller to finish] → [Transcribe all] → [Generate full response] → [Synthesize all] → [Play]

Streaming (fast):

[Transcribe as they speak] → [Start generating on partial transcript] → [Synthesize as tokens arrive] → [Play immediately]

Every component in our stack supports streaming:

Deepgram provides interim transcription results
Claude streams tokens as they’re generated
Chatterbox synthesizes audio incrementally

3. WebRTC

WebRTC (Web Real-Time Communication) is a protocol for real-time audio/video over the internet. It’s what powers video calls in your browser. Key concepts:

Peer-to-peer — Direct connections between participants
SDP (Session Description Protocol) — How peers negotiate connection parameters
ICE (Interactive Connectivity Establishment) — How peers find a path to connect
Tracks — Individual audio or video streams

We use WebRTC to get audio from the phone system (GoToConnect) into our processing pipeline (LiveKit).

4. LiveKit

LiveKit is an open-source platform for real-time audio/video. Think of it as “Zoom/WebRTC infrastructure as a service.” Key concepts:

Rooms — Virtual spaces where participants connect
Participants — Entities in a room (could be users or AI agents)
Tracks — Audio or video streams published by participants
LiveKit Agents SDK — Framework for building AI agents that participate in rooms

In our system:

Each phone call creates a LiveKit room
The WebRTC bridge joins as one participant (representing the caller)
The AI agent joins as another participant
Audio flows between them through the room

5. State Machines

Phone calls have states: ringing, connected, on hold, transferred, ended. Managing these transitions correctly is crucial.

┌─────────┐     ┌───────────┐     ┌────────────┐     ┌─────────┐
│ RINGING │────▶│ CONNECTED │────▶│ CONVERSING │────▶│  ENDED  │
└─────────┘     └───────────┘     └────────────┘     └─────────┘
                      │                  │
                      │                  │
                      ▼                  ▼
               ┌───────────┐     ┌─────────────┐
               │  ON_HOLD  │     │ TRANSFERRING│
               └───────────┘     └─────────────┘

We use Redis to store call state because:

It’s fast (in-memory)
It’s ephemeral (call state doesn’t need to persist forever)
It supports pub/sub (for real-time notifications)

6. Multi-Tenancy

Multiple businesses (tenants) use the same platform. Each tenant has:

Their own phone numbers
Their own AI configuration (personality, knowledge base, tools)
Their own usage tracking and billing
Isolated data (Tenant A can’t see Tenant B’s calls)

This is implemented through:

Tenant IDs on all database records
Scoped API keys
Request-level tenant context

The Technology Stack

Languages & Frameworks

Component	Language	Framework	Why
WebRTC Bridge	Python	aiortc	Best WebRTC library for Python
AI Agent	Python	LiveKit Agents SDK	Native Python SDK
API Gateway	Python	FastAPI	Async, fast, great docs
Background Workers	Python	Celery or native async	Job processing

External Services

Service	Purpose	Why This One
GoToConnect	Phone system	Existing, full API, unlimited plan
LiveKit Cloud	Real-time audio	Industry standard, Agents SDK
Deepgram	Speech-to-text	Low latency, streaming, accurate
Anthropic Claude	AI reasoning	Best reasoning, streaming
RunPod	GPU hosting	Cost-effective A5000
Chatterbox	Text-to-speech	Free, MIT license, paralinguistics

Databases & Storage

Service	Purpose	Why This One
PostgreSQL	Relational data	Tenants, configs, call logs
Redis	Cache & state	Call state machine, sessions
DO Spaces	Object storage	Voice samples, recordings

Infrastructure

Service	Purpose
DigitalOcean	Cloud hosting
Dokploy	Container orchestration
RunPod	GPU instances for TTS

Project Structure

Here’s how the codebase is organized:

voice-by-aiconnected/
├── docs/                          # Documentation (you are here)
│   ├── 00-MASTER-PROJECT-TASK-LIST.md
│   ├── 01-system-architecture.md
│   └── ...
│
├── services/                      # Backend services
│   ├── api-gateway/              # Public API (FastAPI)
│   │   ├── app/
│   │   │   ├── main.py
│   │   │   ├── routers/
│   │   │   ├── models/
│   │   │   └── dependencies/
│   │   ├── Dockerfile
│   │   └── requirements.txt
│   │
│   ├── webrtc-bridge/            # GoToConnect ↔ LiveKit bridge
│   │   ├── bridge/
│   │   │   ├── main.py
│   │   │   ├── goto_client.py
│   │   │   ├── livekit_client.py
│   │   │   └── audio_handler.py
│   │   ├── Dockerfile
│   │   └── requirements.txt
│   │
│   ├── agent-service/            # LiveKit AI agent
│   │   ├── agent/
│   │   │   ├── main.py
│   │   │   ├── pipeline.py       # STT → LLM → TTS
│   │   │   ├── tools.py          # Function calling
│   │   │   └── prompts.py
│   │   ├── Dockerfile
│   │   └── requirements.txt
│   │
│   └── worker-service/           # Background jobs
│       ├── worker/
│       │   ├── main.py
│       │   ├── tasks/
│       │   └── handlers/
│       ├── Dockerfile
│       └── requirements.txt
│
├── shared/                        # Shared code
│   ├── database/
│   │   ├── models.py             # SQLAlchemy models
│   │   └── migrations/           # Alembic migrations
│   ├── schemas/                   # Pydantic schemas
│   ├── utils/
│   └── config/
│
├── integrations/                  # Provider integrations
│   ├── gotoconnect/
│   ├── livekit/
│   ├── deepgram/
│   ├── anthropic/
│   ├── chatterbox/
│   └── n8n/
│
├── tests/
│   ├── unit/
│   ├── integration/
│   └── e2e/
│
├── infrastructure/
│   ├── docker-compose.yml        # Local development
│   ├── docker-compose.prod.yml   # Production
│   └── dokploy/                  # Deployment configs
│
└── scripts/
    ├── setup.sh
    ├── migrate.sh
    └── seed.sh

Key Files You’ll Work With

WebRTC Bridge (`services/webrtc-bridge/`)

This is the trickiest part of the system. It:

Receives calls from GoToConnect via WebRTC
Extracts audio frames
Publishes them to a LiveKit room
Receives AI audio from LiveKit
Sends it back to GoToConnect

# Simplified example of what the bridge does
class WebRTCBridge:
    async def handle_incoming_call(self, call_id: str, sdp_offer: str):
        # Create a LiveKit room for this call
        room = await self.livekit.create_room(f"call-{call_id}")
        
        # Set up WebRTC connection to GoToConnect
        pc = RTCPeerConnection()
        
        @pc.on("track")
        async def on_track(track):
            if track.kind == "audio":
                # Forward caller's audio to LiveKit
                await self.forward_to_livekit(track, room)
        
        # Complete WebRTC handshake
        await pc.setRemoteDescription(RTCSessionDescription(sdp=sdp_offer, type="offer"))
        answer = await pc.createAnswer()
        await pc.setLocalDescription(answer)
        
        return answer.sdp

AI Agent (`services/agent-service/`)

This is where the magic happens. The agent:

Joins a LiveKit room
Listens to the audio track from the bridge
Transcribes it with Deepgram
Generates a response with Claude
Synthesizes speech with Chatterbox
Publishes the audio back to the room

# Simplified example using LiveKit Agents SDK
from livekit.agents import Agent, AgentContext
from livekit.plugins import deepgram, anthropic

class VoiceAgent:
    def __init__(self, tenant_config: TenantConfig):
        self.stt = deepgram.STT()
        self.llm = anthropic.LLM(model="claude-sonnet-4-20250514")
        self.tts = ChatterboxTTS(tenant_config.voice_id)
        self.knowledge_base = tenant_config.knowledge_base
        
    async def run(self, ctx: AgentContext):
        # Build the system prompt with business context
        system_prompt = self.build_prompt(self.knowledge_base)
        
        # Create the conversational pipeline
        pipeline = VoicePipeline(
            stt=self.stt,
            llm=self.llm,
            tts=self.tts,
            system_prompt=system_prompt,
            tools=self.get_tools()
        )
        
        # Run until the call ends
        await pipeline.run(ctx)

Call State Machine (`shared/state/`)

Manages the lifecycle of each call:

from enum import Enum
from redis import Redis

class CallState(Enum):
    RINGING = "ringing"
    CONNECTED = "connected"
    CONVERSING = "conversing"
    ON_HOLD = "on_hold"
    TRANSFERRING = "transferring"
    ENDED = "ended"

class CallStateMachine:
    def __init__(self, redis: Redis, call_id: str):
        self.redis = redis
        self.call_id = call_id
        self.key = f"call:{call_id}:state"
    
    async def transition(self, new_state: CallState) -> bool:
        """Attempt to transition to a new state."""
        current = await self.get_state()
        
        if self.is_valid_transition(current, new_state):
            await self.redis.set(self.key, new_state.value)
            await self.publish_event(current, new_state)
            return True
        return False
    
    def is_valid_transition(self, from_state: CallState, to_state: CallState) -> bool:
        """Check if the transition is allowed."""
        valid_transitions = {
            CallState.RINGING: [CallState.CONNECTED, CallState.ENDED],
            CallState.CONNECTED: [CallState.CONVERSING, CallState.ENDED],
            CallState.CONVERSING: [CallState.ON_HOLD, CallState.TRANSFERRING, CallState.ENDED],
            CallState.ON_HOLD: [CallState.CONVERSING, CallState.ENDED],
            CallState.TRANSFERRING: [CallState.ENDED],
        }
        return to_state in valid_transitions.get(from_state, [])

Development Workflow

Setting Up Your Environment

Clone the repository

git clone <repo-url>
cd voice-by-aiconnected

Copy environment template

cp .env.example .env
# Fill in your API keys

Start local services

docker-compose up -d

Run migrations

./scripts/migrate.sh

Start development servers

# In separate terminals:
cd services/api-gateway && uvicorn app.main:app --reload
cd services/webrtc-bridge && python -m bridge.main
cd services/agent-service && python -m agent.main

Testing a Call Locally

For local development, you’ll use test audio files instead of real phone calls:

# Simulate an inbound call with test audio
python scripts/simulate_call.py --audio test_audio/hello.wav

Running Tests

# Unit tests
pytest tests/unit/

# Integration tests (requires services running)
pytest tests/integration/

# End-to-end tests (requires all services + credentials)
pytest tests/e2e/

Common Patterns You’ll See

1. Async Everything

Almost all our code is async because we’re dealing with I/O-bound operations (network calls, audio streaming). Get comfortable with:

async def process_audio(stream: AsyncIterator[bytes]) -> AsyncIterator[str]:
    async for chunk in stream:
        result = await transcribe(chunk)
        yield result

2. Dependency Injection

We use FastAPI’s dependency injection for database sessions, authentication, tenant context:

@router.post("/calls")
async def create_call(
    request: CreateCallRequest,
    db: Session = Depends(get_db),
    tenant: Tenant = Depends(get_current_tenant)
):
    # tenant is automatically extracted from the API key
    call = await create_call_for_tenant(db, tenant, request)
    return call

3. Event-Driven Communication

Services communicate through events, not direct calls:

# When a call connects
await event_bus.publish("call.connected", {
    "call_id": call_id,
    "tenant_id": tenant_id,
    "timestamp": datetime.utcnow()
})

# Another service listens
@event_handler("call.connected")
async def handle_call_connected(event: Event):
    await start_agent_for_call(event.data["call_id"])

4. Circuit Breakers

External services can fail. We use circuit breakers to fail fast:

from circuitbreaker import circuit

@circuit(failure_threshold=5, recovery_timeout=30)
async def call_deepgram(audio: bytes) -> str:
    # If this fails 5 times, the circuit opens
    # and we fail immediately for 30 seconds
    return await deepgram_client.transcribe(audio)

5. Graceful Degradation

When components fail, we degrade gracefully:

async def synthesize_speech(text: str) -> bytes:
    try:
        # Try Chatterbox first
        return await chatterbox.synthesize(text)
    except ChatterboxError:
        # Fall back to Resemble API
        logger.warning("Chatterbox failed, using fallback")
        return await resemble_api.synthesize(text)
    except Exception:
        # Last resort: pre-recorded message
        return load_audio("fallback_messages/technical_difficulty.wav")

Key Challenges You’ll Face

1. Latency Optimization

Every millisecond matters. You’ll need to:

Profile everything
Avoid blocking operations
Stream wherever possible
Cache aggressively
Minimize network hops

2. Audio Quality

Telephone audio is 8kHz, muddy, and often has background noise. You’ll need to:

Handle different audio formats
Resample correctly
Understand codec differences
Deal with packet loss

3. Conversation Flow

Natural conversations have:

Interruptions (barge-in)
Pauses
Overlapping speech
Misunderstandings

The AI needs to handle all of these gracefully.

4. Error Recovery

Lots of things can fail:

Network issues
API rate limits
Audio dropout
Service outages

Your code needs to handle these without dropping calls.

5. Concurrency

Multiple calls happen simultaneously. You need to:

Avoid race conditions
Manage connection pools
Handle resource contention
Scale horizontally

Glossary

Term	Definition
STT	Speech-to-Text. Converting audio to text.
TTS	Text-to-Speech. Converting text to audio.
LLM	Large Language Model. AI that generates text (like Claude).
WebRTC	Web Real-Time Communication. Protocol for real-time audio/video.
SDP	Session Description Protocol. How WebRTC peers negotiate.
PBX	Private Branch Exchange. A business phone system.
PSTN	Public Switched Telephone Network. The traditional phone system.
SIP	Session Initiation Protocol. Protocol for VoIP calls.
VAD	Voice Activity Detection. Detecting when someone is speaking.
TTFB	Time to First Byte. How long until the first response arrives.
RTF	Real-Time Factor. If RTF < 1, synthesis is faster than real-time.
Barge-in	When the caller interrupts the AI while it’s speaking.
Warm Transfer	Transferring a call after briefing the recipient.
Blind Transfer	Transferring a call without briefing the recipient.
Multi-tenant	One system serving multiple separate customers.

How to Get Help

Documentation

This document — Start here for overview
Master Project Task List — Overall project plan
Individual specification documents — Deep dives into each component
Skills folder — API reference for each provider

Code

Read the tests — Tests show how things are supposed to work
Read the types — Type hints document expected inputs/outputs
Read the docstrings — Functions should explain what they do

People

Ask questions — No question is too basic
Review PRs — See how others solve problems
Pair program — Learn by doing together

Your First Tasks

If you’re new to the project, here are good starting points:

Beginner

Set up your local development environment
Run the test suite and make sure everything passes
Read through the API gateway routes to understand the API surface
Add a simple new endpoint (e.g., health check with more details)

Intermediate

Add a new tool that the AI can call (e.g., check business hours)
Improve error messages in a service
Add metrics/logging to an existing component
Write integration tests for an existing feature

Advanced

Implement a new call feature (e.g., call recording)
Optimize latency in the voice pipeline
Add a new STT/TTS provider as a fallback
Implement a complex state machine transition

Final Thoughts

Voice AI is a fascinating intersection of several technologies:

Real-time systems
AI/ML
Telecommunications
Distributed systems

It’s challenging because everything happens in real-time. You can’t hide latency behind a loading spinner. You can’t ask the user to refresh the page. The conversation has to flow naturally, and if anything goes wrong, it’s immediately obvious. But when it works, it’s magical. A computer that you can talk to like a person, that understands you, that helps you — that’s the future we’re building. Welcome to the team.

Last updated: 2026-01-16

Overview

aiConnected OS

Business Platform

Apps & Modules

Neurigraph

Acquired Intelligence

Spatial Computing

Papers & Research

Supporting Docs

Archive

​Voice by aiConnected — A Developer’s Introduction

​What This Document Is

​What We’re Building

​The One-Sentence Version

​The One-Paragraph Version

​The Visual Version

​Core Concepts You Need to Understand

​1. The Voice Pipeline

​2. Streaming vs. Batch Processing

​3. WebRTC

​4. LiveKit

​5. State Machines

​6. Multi-Tenancy

​The Technology Stack

​Languages & Frameworks

​External Services

​Databases & Storage

​Infrastructure

​Project Structure

​Key Files You’ll Work With

​WebRTC Bridge (services/webrtc-bridge/)

​AI Agent (services/agent-service/)

​Call State Machine (shared/state/)

​Development Workflow

​Setting Up Your Environment

​Testing a Call Locally

​Running Tests

​Common Patterns You’ll See

​1. Async Everything

​2. Dependency Injection

​3. Event-Driven Communication

​4. Circuit Breakers

​5. Graceful Degradation

​Key Challenges You’ll Face

​1. Latency Optimization

​2. Audio Quality

​3. Conversation Flow

​4. Error Recovery

​5. Concurrency

​Glossary

​How to Get Help

​Documentation

​Code

​People

​Your First Tasks

​Beginner

​Intermediate

​Advanced

​Final Thoughts