Demo: Real-time Voice Agents with Video Avatar

An AI-powered video avatar that engages visitors in real-time voice conversations at AWS events. Supports content overlays (architecture diagrams, use cases) via tool calls.

Prerequisites

Requirement	Version	Notes
Node.js	18+	Required for both demos
npm	9+	Comes with Node.js
Python	3.11+	Pipecat demo only
AWS CLI	2.x	For deployment and Bedrock access
AWS credentials	—	Configured via `aws configure` or env vars

API keys from AWS Partners (also available on AWS Marketplace):

Key	Required for	AWS Partner
`TAVUS_API_KEY`	Both demos	Tavus
`DEEPGRAM_API_KEY`	Pipecat demo	Deepgram
`CARTESIA_API_KEY`	Pipecat demo	Cartesia
`DAILY_API_KEY`	Pipecat cloud deploy	Daily

Quick Start

# Tavus Avatar (managed pipeline)
cd tavus-avatar && cp env.example .env.local
# Edit .env.local with your TAVUS_API_KEY and TAVUS_PERSONA_ID
npm install && npm run dev:all

# Pipecat (self-hosted pipeline)
cd tavus-pipecat-example && cp .env.example .env
# Edit .env with your API keys
pip install -r requirements.txt && npm run dev:all

Open http://localhost:3000.

Choose Your Implementation

This repo contains two independent implementations of the same demo:

	`tavus-avatar/`	`tavus-pipecat-example/`
Approach	Tavus hosted CVI¹	Self-hosted Pipecat pipeline
Frontend	Next.js + Daily.js	React + WebRTC
Backend	Next.js API routes → Tavus API	Python Pipecat server
Pipeline control	Tavus orchestrates STT²/LLM³/TTS⁴	You orchestrate everything in Python
Pipeline modes	Single (Tavus-managed cascaded)	Cascaded + Amazon Nova Sonic toggle
When to choose	Quick demos, minimal setup, managed infrastructure	Custom pipelines, multi-model support, full control
Best for	Prototyping, simple use cases	Production, complex workflows, experimentation

Both share the same tool definitions and content overlays. System prompts are isolated per event in prompts/{EVENT_CONFIG}/.

graph LR
    subgraph "Tavus Avatar"
        TA[Browser] --> TB[Daily.js] --> TC[Tavus API] --> TD[Managed Pipeline]
    end
    
    subgraph "Pipecat Demo"
        PA[Browser] --> PB[WebRTC] --> PC[Python Server] --> PD[Custom Pipeline]
        PD --> PE[Deepgram STT]
        PD --> PF[Bedrock LLM]
        PD --> PG[Cartesia TTS]
        PD --> PH[Nova Sonic]
    end

Run Tavus Avatar Demo

A Next.js app that delegates pipeline orchestration to Tavus. You configure the persona, STT²/LLM³/TTS⁴ engines, and tools via the Tavus API. Daily.js handles the WebRTC transport.

Prerequisites

Node.js 18+ and npm
A Tavus account with an API key and configured persona

Set Up Environment

cd tavus-avatar
npm install
cp env.example .env.local

Example .env.local:

TAVUS_API_KEY=your_tavus_api_key_here
TAVUS_PERSONA_ID=your_persona_id_here

Start Development Server

# All-in-one command
npm run dev:all

# Or step by step
npm run dev
curl -X POST http://localhost:3000/api/persona/setup-tools

Open http://localhost:3000.

Validate Environment

# Check API key and persona
curl -H "x-api-key: $TAVUS_API_KEY" \
  "https://tavusapi.com/v2/personas/$TAVUS_PERSONA_ID"

Pipeline Architecture

Uses a cascaded pipeline orchestrated by Pipecat inside Tavus:

Audio In → Noise Cancellation (local) → VAD⁵ → STT² → LLM³ → TTS⁴ → Audio Out

Layer	Engine	Role
Noise Cancellation	Krisp (local macOS)	Filters background noise via a virtual microphone device
VAD⁵	Tavus Sparrow	Turn-taking and interruption detection
STT²	`tavus-advanced` (via Tavus)	Real-time speech-to-text
LLM³	`tavus-gpt-oss` (via Tavus)	Reasoning and response generation
TTS⁴	`sonic-3` (Cartesia via Tavus)	Natural-sounding speech synthesis

STT²/TTS⁴ engines are managed within Tavus. See Update Tavus Persona for details.

Run Pipecat Demo

A Python Pipecat server with a React frontend. You control the full pipeline in code. Supports two pipeline modes that visitors can toggle at the kiosk:

Cascaded: Deepgram STT² → Bedrock Claude → Cartesia TTS⁴ (best-of-breed at each layer)
Nova Sonic: Amazon Nova 2 Sonic speech-to-speech on Bedrock (single model, 15 languages)

Both modes use Tavus for avatar rendering and Pipecat for orchestration.

Prerequisites

Python 3.11+
Node.js 18+ and npm
API keys: Deepgram, Cartesia, Tavus, AWS credentials (for Bedrock)

Set Up Environment

cd tavus-pipecat-example

# Backend
python -m venv env && source env/bin/activate
pip install -r requirements.txt
cp .env.example .env

Example .env:

DEEPGRAM_API_KEY=your_deepgram_key_here
CARTESIA_API_KEY=your_cartesia_key_here
TAVUS_API_KEY=your_tavus_key_here
TAVUS_REPLICA_ID=your_replica_id_here
TAVUS_REPLICA_ID_NOVA_SONIC=your_nova_sonic_replica_id_here
AWS_REGION_NOVA_SONIC=ap-south-1
NOVA_SONIC_VOICE_ID=matthew
CARTESIA_VOICE_ID=79a125e8-cd45-4c13-8a67-188112f4dd22
AWS_ACCESS_KEY_ID=your_aws_key_here
AWS_SECRET_ACCESS_KEY=your_aws_secret_here

Start Development Server

# All-in-one command
npm run dev:all

# Or step by step
python tavus-pipecat.py --transport webrtc --host 0.0.0.0 --port 7860 &
cd frontend && npm start

Open http://localhost:3000. Click Start demo, choose a pipeline mode, and begin.

Validate Environment

# Test API keys
curl -H "Authorization: Token $DEEPGRAM_API_KEY" \
  "https://api.deepgram.com/v1/projects"

curl -H "X-API-Key: $CARTESIA_API_KEY" \
  "https://api.cartesia.ai/voices"

aws bedrock list-foundation-models --region $AWS_REGION_NOVA_SONIC

Pipeline Architecture

Cascaded mode:

Audio In -> Deepgram Nova 3 STT -> Bedrock Claude LLM -> Cartesia TTS -> Tavus Avatar -> Audio/Video Out

Nova Sonic mode:

Audio In -> Amazon Nova 2 Sonic (STT+LLM+TTS) -> Tavus Avatar -> Audio/Video Out

Environment Variables

Variable	Description
`DEEPGRAM_API_KEY`	Deepgram API key (cascaded STT²)
`CARTESIA_API_KEY`	Cartesia API key (cascaded TTS⁴)
`TAVUS_API_KEY`	Tavus API key (avatar)
`TAVUS_REPLICA_ID`	Tavus replica ID for cascaded mode (see Available Avatars)
`TAVUS_REPLICA_ID_NOVA_SONIC`	Tavus replica ID for Nova Sonic mode — use a different avatar to visually distinguish the two modes
`DAILY_API_KEY`	Daily.co API key (required for cloud deployment with Daily transport)
`REACT_APP_TRANSPORT`	WebRTC transport (`daily` or `webrtc`). Set automatically by CI/CD — not user-configured.
`AWS_REGION_NOVA_SONIC`	AWS region for Nova Sonic (default: `ap-south-1`)
`NOVA_SONIC_VOICE_ID`	Nova Sonic voice ID (default: `matthew`; use `arjun` for Hindi)
`CARTESIA_VOICE_ID`	Cartesia voice UUID (default: `79a125e8-...` British Lady)
`AWS_ACCESS_KEY_ID`	AWS credentials (optional if using default credential chain)
`AWS_SECRET_ACCESS_KEY`	AWS credentials (optional if using default credential chain)

Deploying to AWS

The Pipecat demo deploys to AWS with a GitHub Actions CI/CD pipeline that triggers only on changes to tavus-pipecat-example/ or prompts/.

Architecture:

Backend: ECS Fargate (Docker container) behind an ALB
Frontend: S3 + CloudFront (static React build)
CloudFront: Serves frontend and proxies /api/* to the ALB (single HTTPS domain)
Secrets: SSM Parameter Store (SecureString)
Transport: Daily.js for cloud (reliable WebRTC relay), SmallWebRTC for local dev

graph TD
    User([User / Browser])

    User -->|HTTPS| CF[CloudFront]

    subgraph AWS Cloud
        CF -->|Static assets| S3[S3 Bucket\nReact Frontend]
        CF -->|/api/* and /start| ALB[Application Load Balancer]
        ALB --> ECS[ECS Fargate\nPipecat Bot Container]

        SSM[SSM Parameter Store\nSecureString secrets] -->|Secrets at startup| ECS
        ECR[ECR\nContainer Registry] -->|Pull image| ECS
    end

    GHA[GitHub Actions\nCI/CD] -->|Push image| ECR
    GHA -->|Sync build| S3

    ECS -->|WebRTC relay| Daily[Daily.co]
    ECS -->|Speech-to-text| Deepgram[Deepgram STT]
    ECS -->|LLM inference| Bedrock[Bedrock Claude\nLLM]
    ECS -->|Text-to-speech| Cartesia[Cartesia TTS]
    ECS -->|Avatar rendering| Tavus[Tavus Avatar]
    ECS -->|Speech-to-speech| NovaSonic[Bedrock Nova Sonic\nSpeech-to-Speech]

1. Store API keys in SSM Parameter Store

aws ssm put-parameter --name /tavus-pipecat/DEEPGRAM_API_KEY --type SecureString --value "your-key"
aws ssm put-parameter --name /tavus-pipecat/CARTESIA_API_KEY --type SecureString --value "your-key"
aws ssm put-parameter --name /tavus-pipecat/TAVUS_API_KEY --type SecureString --value "your-key"
aws ssm put-parameter --name /tavus-pipecat/TAVUS_REPLICA_ID --type SecureString --value "your-replica-id"
aws ssm put-parameter --name /tavus-pipecat/TAVUS_REPLICA_ID_NOVA_SONIC --type SecureString --value "your-replica-id"
aws ssm put-parameter --name /tavus-pipecat/DAILY_API_KEY --type SecureString --value "your-daily-api-key"

2. Deploy infrastructure (one-time)

aws cloudformation deploy \
  --template-file tavus-pipecat-example/infra/template.yaml \
  --stack-name tavus-pipecat \
  --parameter-overrides \
    VpcId=vpc-xxx \
    PublicSubnetIds=subnet-aaa,subnet-bbb \
  --capabilities CAPABILITY_NAMED_IAM

Note the outputs — you'll need them for GitHub Actions secrets.

3. Set up GitHub Actions OIDC

Create a GitHub OIDC identity provider and IAM role so GitHub Actions can assume credentials without static keys:

# Create OIDC provider (one-time per account)
aws iam create-open-id-connect-provider \
  --url https://token.actions.githubusercontent.com \
  --client-id-list sts.amazonaws.com \
  --thumbprint-list 6938fd4d98bab03faadb97b34396831e3780aea1

# Create IAM role with trust policy for your repo, then attach
# ECR, ECS, S3, and CloudFront permissions (see infra/template.yaml)

4. Configure GitHub Actions secrets

Secret	Value (from CloudFormation outputs)
`AWS_ROLE_ARN`	IAM role ARN for GitHub Actions OIDC
`S3_BUCKET_NAME`	`FrontendBucketName` output
`CLOUDFRONT_DISTRIBUTION_ID`	`CloudFrontDistributionId` output

Also set these repository variables (Settings → Variables → Actions) to configure the event and voices per deployment:

Variable	Example (Mumbai)	Example (Sydney)
`AWS_REGION`	`ap-south-1`	`ap-southeast-2`
`EVENT_CONFIG`	`aws-summit-mumbai-2026`	`aws-summit-sydney-2026`
`NOVA_SONIC_VOICE_ID`	`matthew`	`matthew`
`CARTESIA_VOICE_ID`	`79a125e8-cd45-4c13-8a67-188112f4dd22`	`79a125e8-cd45-4c13-8a67-188112f4dd22`

5. Push to deploy

The workflow at .github/workflows/deploy-pipecat.yml runs automatically on push to main when files in tavus-pipecat-example/ or prompts/ change. It:

Builds the Docker image and pushes to ECR
Forces a new ECS Fargate deployment and waits for stability
Builds the React frontend (with REACT_APP_API_URL='' for same-origin CloudFront)
Syncs to S3 and invalidates the CloudFront cache

The frontend and backend deploy in parallel. Access the demo at the CloudFront domain (CloudFrontDomainName output).

Note: Regenerate frontend/package-lock.json with Node 20 (fnm use 20 && npm install) to match CI. The first CloudFormation deploy is slow (~15 min) as the ECS service waits for a Docker image — subsequent deploys via CI/CD are faster.

Browser Compatibility

Supported Browsers:

Chrome 88+ (recommended)
Firefox 85+
Safari 14.1+
Edge 88+

WebRTC Requirements:

Secure context (HTTPS or localhost)
Microphone permissions
WebRTC support (enabled by default in modern browsers)

Known Issues:

Safari may require user gesture before audio playback
Firefox requires media.navigator.permission.disabled=true for localhost development

Performance Considerations

Latency Optimization:

Tavus Avatar: ~200-400ms end-to-end latency (managed pipeline)
Pipecat Cascaded: ~300-500ms (Deepgram + Bedrock + Cartesia)
Nova Sonic: ~400-600ms (single model, higher quality)

Concurrent Users:

Local Development: 1-2 concurrent sessions
AWS Deployment: 10-50 concurrent sessions (auto-scaling ECS)
Resource Usage: ~1GB RAM, 0.5 CPU per active session

Optimization Tips:

Use Daily.js transport for production (better WebRTC handling)
Enable Krisp noise cancellation for cleaner audio input
Monitor ECS CPU/memory metrics for scaling decisions

Security

API Key Management:

Store keys in environment variables, never in code
Use AWS SSM Parameter Store (SecureString) for production
Rotate keys regularly and monitor usage

CORS Configuration:

// Next.js API routes
const corsHeaders = {
  'Access-Control-Allow-Origin': process.env.NODE_ENV === 'production' 
    ? 'https://your-domain.com' 
    : '*',
  'Access-Control-Allow-Methods': 'GET, POST, OPTIONS',
  'Access-Control-Allow-Headers': 'Content-Type, Authorization'
}

Rate Limiting:

Implement per-IP session limits
Use AWS WAF for DDoS protection
Monitor API usage and set billing alerts

Troubleshooting

Microphone Permissions

Chrome/Edge:

Click the microphone icon in the address bar
Select "Always allow" for the site
Refresh the page

Firefox:

Click the shield icon → Permissions → Microphone
Select "Allow" and check "Remember this decision"

Safari:

Safari → Settings → Websites → Microphone
Set your site to "Allow"

API Key Validation

# Tavus
curl -H "x-api-key: $TAVUS_API_KEY" "https://tavusapi.com/v2/personas"

# Deepgram
curl -H "Authorization: Token $DEEPGRAM_API_KEY" "https://api.deepgram.com/v1/projects"

# Cartesia
curl -H "X-API-Key: $CARTESIA_API_KEY" "https://api.cartesia.ai/voices"

# AWS Bedrock
aws bedrock list-foundation-models --region us-east-1

WebRTC Connection Issues

Symptoms: No audio/video, connection timeouts Solutions:

Check browser console for WebRTC errors
Verify HTTPS/localhost secure context
Test with different networks (corporate firewalls may block WebRTC)
Use Daily.js transport for better NAT traversal

CORS Problems

Symptoms: Network errors, blocked requests Solutions:

Check browser console for CORS errors
Verify API endpoints allow your domain
For local development, use --disable-web-security flag (Chrome only)
Ensure proper CORS headers in API responses

Monitoring & Debugging

Logging:

# Pipecat server logs
python tavus-pipecat.py --log-level DEBUG

# Browser console
# Check Network tab for API calls
# Check Console for WebRTC errors

Health Checks:

# Tavus Avatar
curl http://localhost:3000/api/health

# Pipecat
curl http://localhost:7860/health

AWS CloudWatch:

ECS service metrics (CPU, memory, task count)
ALB target health and response times
CloudFront cache hit rates and errors

Advanced Topics

Update Tavus Persona

For tavus-avatar only. The persona's system prompt, tools, and engines are managed via the Tavus API.

Update the system prompt

# Set EVENT_CONFIG to match your event (e.g. aws-summit-mumbai-2026)
SYSTEM_PROMPT=$(python3 -c "import json; print(json.dumps(open('prompts/$EVENT_CONFIG/system-instruction.md').read()))")

curl -X PATCH "https://tavusapi.com/v2/personas/$TAVUS_PERSONA_ID" \
  -H "x-api-key: $TAVUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d "[{\"op\": \"replace\", \"path\": \"/system_prompt\", \"value\": $SYSTEM_PROMPT}]"

Sync tool definitions

curl -X POST "http://localhost:3000/api/persona/setup-tools"

Change STT²/TTS⁴ engines

curl -X PATCH "https://tavusapi.com/v2/personas/$TAVUS_PERSONA_ID" \
  -H "x-api-key: $TAVUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[{"op": "replace", "path": "/layers/stt/stt_engine", "value": "tavus-advanced"}]'

Electron Kiosk Shell

The agent-kiosk-shell/ directory contains a generic Electron wrapper for running either demo in a borderless full-screen kiosk window. It appends autostart=1&shell=electron so sessions start immediately.

cd agent-kiosk-shell
npm install
cp .env.example .env

# Local dev
npm start -- --target-url=http://localhost:3000

# Deployed app
npm start -- --target-url=https://your-app.vercel.app --display-number=2

Keyboard shortcuts:

Cmd+R — Reload
Cmd+B — Show disconnected screen
Cmd+Q — Quit

Noise Cancellation

Noise cancellation runs locally on the kiosk Mac using Krisp, not inside the pipeline:

Install the Krisp desktop app.
Enable Noise Cancellation for the microphone.
In macOS System Settings → Sound → Input, select Krisp Microphone.
The browser uses the Krisp virtual mic automatically.

Shared Resources

System Prompt

Both implementations load the system prompt from prompts/{EVENT_CONFIG}/system-instruction.md. Set EVENT_CONFIG to the event directory name (e.g. aws-summit-mumbai-2026) to switch event context — prompts, greetings, voice IDs, and content items are all isolated per event. Copy prompts/_template/ to add a new event.

Tool Calls

Both implementations support three tools:

show_content — Display content overlays (architecture diagram, voice AI overview, use cases)
show_schedule — Show summit schedule as markdown tables
dismiss_content — Return to full-screen avatar view

Content Pages

Static HTML content pages are served from tavus-pipecat-example/frontend/public/content/ and tavus-avatar/src/app/content/.

Usage

Click Start demo on the landing page.
Grant microphone permissions when prompted.
Speak with the AI avatar in real-time.
Ask the avatar to show content (e.g., "show me the architecture diagram").
Press End or Escape to end the session.

Keyboard shortcuts:

Ctrl+D — Toggle microphone mute
Ctrl+F — Cycle microphone devices (tavus-avatar only)
Ctrl+G — Cycle speaker devices (tavus-avatar only)

Available Tavus Avatars

Tavus provides stock replicas usable without custom recordings. Set TAVUS_REPLICA_ID and TAVUS_REPLICA_ID_NOVA_SONIC in SSM Parameter Store to change the avatar per deployment.

Female replicas:

Name	Replica ID	Setting
Olivia	`rc2146c13e81`	Neutral background
Olivia - Office	`r9fa0878977a`	Office setting
Anna	`r6ae5b6efc9d`	Neutral background
Anna - Office	`r4dcf31b60e1`	Office setting
Julia	`rdc96ac37313`	Neutral background
Julia - Home Office	`rb43357fb2ee`	Home office
Gloria - Conversational	`rbe2c395e725`	Conversational style
Rose	`r1af76e94d00`	Neutral background
Luna	`r9d30b0e55ac`	Neutral background
Helen - Casual	`r12d3eb75ec2`	Casual setting
Lucy - Studio	`r53a461095cf`	Studio setting
Samantha - Office V2	`r38e4c3bc562`	Office setting
Gabby	`rdf61be0d4e1`	Neutral background
Ruby	`re3a705cf66a`	Neutral background
Destiny	`r38a383b0173`	Neutral background
Celine - Studio	`r4f5b5ef55c8`	Studio setting

Male replicas:

Name	Replica ID	Setting
Charlie	`rf4703150052`	Neutral background
Benjamin	`r1a4e22fa0d9`	Neutral background
Raj	`ra066ab28864`	Neutral background
Raj - Doctor	`r18e9aebdc33`	Doctor setting
Daniel - Office	`r72f7f7f7c8b`	Office setting
Owen	`r9458111c64a`	Neutral background
Carter	`rca8a38779a8`	Neutral background
Liam	`r90a0339d496`	Neutral background
Diego - Office V2	`r044d76f4490`	Office setting
Danny	`r62baeccd777`	Neutral background
James	`r92debe21318`	Neutral background
Nathan - Bookshelf	`rfe12d8b9597`	Bookshelf setting
Lucas - Studio	`r5f0577fc829`	Studio setting
Victor - Casual	`r1d7cf9edbb4`	Casual setting
Kai	`r31e11adf1d3`	Neutral background
Zane	`r24efb3b9bef`	Neutral background

To update the avatar for a deployment:

# Example: switch Sydney Nova Sonic to Olivia - Office
aws ssm put-parameter \
  --name /tavus-pipecat/TAVUS_REPLICA_ID_NOVA_SONIC \
  --type SecureString --value "r9fa0878977a" --overwrite \
  --region ap-southeast-2

Current per-deployment config:

Deployment	Pipeline	Avatar	Voice
Sydney (`ap-southeast-2`)	Cascaded	Olivia (`rc2146c13e81`)	Cartesia British Lady
Sydney (`ap-southeast-2`)	Nova Sonic	Olivia - Office (`r9fa0878977a`)	olivia (Australian English)
Bengaluru (`ap-south-1`)	Cascaded	(set in SSM)	Cartesia Hindi
Bengaluru (`ap-south-1`)	Nova Sonic	(set in SSM)	arjun (Hindi)

Available Nova Sonic voices: matthew (polyglot), tiffany (polyglot), amy (British English), olivia (Australian English), ambre (French), beatrice (Italian), arjun (Hindi), and others. See Amazon Nova Sonic docs for the full list.

Guidance for Voice Agents

This demo promotes the Guidance for Voice Agents on AWS reference architecture. Ask the avatar "show me the voice agent guidance" to display it.

Glossary

CVI¹ — Conversational Video Interface: Real-time video avatar with voice interaction
STT² — Speech-to-Text: Converts spoken audio to text transcription
LLM³ — Large Language Model: AI model for understanding and generating text responses
TTS⁴ — Text-to-Speech: Converts text responses to natural-sounding speech
VAD⁵ — Voice Activity Detection: Detects when user starts/stops speaking for turn-taking

Project Structure

prompts/                          # Shared system prompts and knowledge base documents
agent-kiosk-shell/                # Generic Electron kiosk wrapper
tavus-avatar/                     # Tavus hosted CVI (Next.js)
  src/
    app/
      api/conversation/           # Create/end Tavus conversation sessions
      api/persona/setup-tools/    # Patch persona tool definitions
      content/                    # Embeddable overlay pages
    components/                   # React components
    lib/                          # Tavus API client, persona/tool configs
    types/                        # TypeScript type definitions
tavus-pipecat-example/            # Self-hosted Pipecat pipeline (Python + React)
  tavus-pipecat.py                # Pipecat bot (dual pipeline: cascaded + Nova Sonic)
  requirements.txt                # Python dependencies
  frontend/                       # React frontend
    src/
      components/                 # VideoConversation UI
      webrtc-client.js            # WebRTC client
    public/
      content/                    # Static content overlay pages

Future Improvements

See IMPROVEMENTS.md for tracked enhancements.

Tech Stack

tavus-avatar: Next.js 16 / React 19 / TypeScript / Tailwind CSS v4 / Daily.js / Tavus API
tavus-pipecat-example: Python / Pipecat / Deepgram / Cartesia / Amazon Bedrock / Amazon Nova Sonic / Tavus / React
Shared: Tavus avatar rendering, Pipecat orchestration, AWS infrastructure

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
agent-kiosk-shell		agent-kiosk-shell
prompts		prompts
tavus-avatar		tavus-avatar
tavus-pipecat-example		tavus-pipecat-example
.ash-ignore		.ash-ignore
.gitignore		.gitignore
CLAUDE.md		CLAUDE.md
DEMO-GUIDE.md		DEMO-GUIDE.md
EQUIPMENT.md		EQUIPMENT.md
IMPROVEMENTS.md		IMPROVEMENTS.md
README.md		README.md

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

Demo: Real-time Voice Agents with Video Avatar

Prerequisites

Quick Start

Choose Your Implementation

Run Tavus Avatar Demo

Prerequisites

Set Up Environment

Start Development Server

Validate Environment

Pipeline Architecture

Run Pipecat Demo

Prerequisites

Set Up Environment

Start Development Server

Validate Environment

Pipeline Architecture

Environment Variables

Deploying to AWS

1. Store API keys in SSM Parameter Store

2. Deploy infrastructure (one-time)

3. Set up GitHub Actions OIDC

4. Configure GitHub Actions secrets

5. Push to deploy

Browser Compatibility

Performance Considerations

Security

Troubleshooting

Microphone Permissions

API Key Validation

WebRTC Connection Issues

CORS Problems

Monitoring & Debugging

Advanced Topics

Update Tavus Persona

Update the system prompt

Sync tool definitions

Change STT²/TTS⁴ engines

Electron Kiosk Shell

Noise Cancellation

Shared Resources

System Prompt

Tool Calls

Content Pages

Usage

Available Tavus Avatars

Guidance for Voice Agents

Glossary

Project Structure

Future Improvements

Tech Stack

About

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Used by

Contributors

Languages