TTS On-Premises

Inworld TTS On-Premises lets organizations run high-quality text-to-speech models locally — without sending text or audio data to the cloud. It’s built for enterprises that require strict data control, low latency, and compliance with internal or regulatory standards. Inworld TTS On-Premises is available for both the Realtime TTS 1.5 Mini and Realtime TTS 1.5 Max models.

To get started with TTS On-Premises, contact sales@inworld.ai for pricing and access to the container registry.

Why TTS On-Premises

Data stays in your environment

No outbound data transfer. Full ownership of text and audio.

Low-latency, real-time speech

Optimized for production workloads and interactive applications.

Designed for regulated industries

Suitable for air-gapped, private, and compliance-sensitive deployments.

Enterprise-ready deployment

Containerized architecture designed for operational stability.

How it works

Inworld TTS On-Premises is delivered as a GPU-accelerated, Docker-containerized version of the Inworld TTS API. It exposes both REST and gRPC APIs for easy integration.

Port	Protocol	Description
8081	HTTP	REST API (recommended)
9030	gRPC	For gRPC clients

Performance

Latency: Real-time streaming on supported NVIDIA GPUs
Throughput: Multiple concurrent sessions are supported depending on the GPU being utilized

Contact sales@inworld.ai to get a detailed performance report for your specific hardware.

System requirements

Inworld TTS supports all modern cloud NVIDIA GPUs: A100s, H100s, H200, B200, B300. If you have a specific target hardware platform not on this list, please reach out for custom support. The minimum inference machine requirements are as follows:

Component	Requirement
GPU	NVIDIA H100 SXM5 (80GB)
RAM	64GB+ system memory
CPU	8+ cores
Disk	50GB free space
OS	Ubuntu 22.04 LTS
Software	Docker + NVIDIA Container Toolkit
Software	Google Cloud SDK (gcloud CLI)
CUDA	13.0+

Prerequisites

Before deploying TTS On-Premises, ensure the following software is installed on your Ubuntu 22.04 LTS machine.

NVIDIA drivers

Install the latest NVIDIA drivers for your GPU. Follow the official guide at nvidia.com/drivers, or use the following commands on Ubuntu:

# Update packages
sudo apt-get update

# Install basic toolchain and kernel headers
sudo apt-get install -y gcc make wget linux-headers-$(uname -r)

# Install NVIDIA driver (check https://www.nvidia.com/en-us/drivers for the latest version)
sudo apt-get install -y nvidia-driver-580

Docker

Install Docker Engine by following the official guide: Install Docker Engine on Ubuntu. Optionally, add the current user to the docker group so you can run Docker without sudo: Linux post-installation steps.

NVIDIA Container Toolkit

Install the NVIDIA Container Toolkit to enable GPU access from Docker containers. Follow both the Installation and Configuration sections of the official guide: NVIDIA Container Toolkit install guide.

Google Cloud SDK

Install the gcloud CLI by following the official guide: Install the gcloud CLI.

Verify prerequisites

Run the following command to verify that Docker, NVIDIA drivers, and the NVIDIA Container Toolkit are all correctly installed:

docker run --rm --gpus all nvidia/cuda:13.0.0-base-ubuntu22.04 nvidia-smi

You should see your GPU listed in the output alongside the driver version and CUDA version. If this command succeeds, your environment is ready for TTS On-Premises deployment.

Firewall requirements

The TTS On-Premises container listens on the following ports for inbound traffic:

Port	Protocol	Description
8081	HTTP	REST API
9030	gRPC	gRPC API

You will also need to allow the following outbound traffic:

us-central1-docker.pkg.dev on port 443 — GCP Artifact Registry for pulling container images

Quick start

1. Create a GCP service account

Create a service account in your GCP project and generate a key file:

# Create the service account
gcloud iam service-accounts create inworld-tts-onprem \
  --project=<YOUR_GCP_PROJECT> \
  --display-name="Inworld TTS On-Prem" \
  --description="Service account for Inworld TTS on-prem container"

# Create a key file
gcloud iam service-accounts keys create service-account-key.json \
  --iam-account=inworld-tts-onprem@<YOUR_GCP_PROJECT>.iam.gserviceaccount.com \
  --project=<YOUR_GCP_PROJECT>

Send the service account email (e.g., inworld-tts-onprem@<YOUR_GCP_PROJECT>.iam.gserviceaccount.com) to your Inworld contact. Inworld will provide your Customer ID.

3. Authenticate to the container registry

gcloud auth activate-service-account \
  --key-file=service-account-key.json

gcloud auth configure-docker us-central1-docker.pkg.dev

For more authentication options, see Configure authentication to Artifact Registry for Docker.

4. Configure

cp onprem.env.example onprem.env

Edit onprem.env with your values:

INWORLD_CUSTOMER_ID=<your-customer-id>
TTS_IMAGE=us-central1-docker.pkg.dev/inworld-ai-registry/tts-onprem/tts-1.5-mini-h100-onprem:<version>
KEY_FILE=./service-account-key.json

5. Start

./run.sh

The script will:

Check prerequisites (Docker, GPU, NVIDIA Container Toolkit)
Validate your configuration
Fix key file permissions if needed
Pull the Docker image
Start the container
Wait for services to be ready (~3 minutes)

The ML model takes approximately 3 minutes to load on first startup. This is normal.

6. Verify the deployment

Check that the container is running and services are healthy:

./run.sh status

7. Send a test request

curl -X POST http://localhost:8081/tts/v1/voice \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Hello, this is a test of the on-premises TTS system.",
    "voice_id": "Craig",
    "model_id": "inworld-tts-1.5-mini",
    "audio_config": {
      "audio_encoding": "LINEAR16",
      "sample_rate_hertz": 48000
    }
  }'

List available voices

curl http://localhost:8081/tts/v1/voices

For the full API specification, see the Synthesize Speech API reference.

Lifecycle commands

./run.sh              # Start the container
./run.sh stop         # Stop and remove the container
./run.sh status       # Check container and service health
./run.sh logs         # Show recent logs from all services
./run.sh logs -f      # Tail all service logs live
./run.sh logs export  # Export all logs to a timestamped folder
./run.sh restart      # Restart the container

Available images

Image	Model	GPU
`tts-1.5-mini-h100-onprem`	Realtime TTS 1.5 Mini	H100
`tts-1.5-max-h100-onprem`	Realtime TTS 1.5 Max	H100

Registry: us-central1-docker.pkg.dev/inworld-ai-registry/tts-onprem/

Configuration

onprem.env

Variable	Required	Description
`INWORLD_CUSTOMER_ID`	Yes	Your customer ID
`TTS_IMAGE`	Yes	Docker image URL (see Available Images)
`KEY_FILE`	Yes	Path to your GCP service account key file

Logs

# Show recent logs from all services (last 20 lines each)
./run.sh logs

# Tail all service logs live
./run.sh logs -f

# Export all logs to a timestamped folder
./run.sh logs export

Individual service logs:

docker exec inworld-tts-onprem tail -f /var/log/tts-v3-trtllm.log        # ML server
docker exec inworld-tts-onprem tail -f /var/log/tts-normalization.log     # Text normalization
docker exec inworld-tts-onprem tail -f /var/log/public-tts-service.log    # TTS service
docker exec inworld-tts-onprem tail -f /var/log/grpc-gateway.log          # HTTP gateway
docker exec inworld-tts-onprem tail -f /var/log/w-proxy.log               # gRPC proxy
docker exec inworld-tts-onprem tail -f /var/log/supervisord.log           # Supervisor

Troubleshooting

Issue	Solution
”INWORLD_CUSTOMER_ID is required”	Set `INWORLD_CUSTOMER_ID` in `onprem.env`
”GCP credentials file not found”	Check that `KEY_FILE` in `onprem.env` points to a valid file
”Credentials file is not readable”	Fix permissions on host: `chmod 644 <your-key-file>.json`
”Topic not found”	Verify your `INWORLD_CUSTOMER_ID` matches the PubSub topic name
”Permission denied for topic”	Ensure Inworld has granted your service account publish access
Slow startup (~3 min)	Normal — text processing grammars take time to initialize

# Check service status
docker exec inworld-tts-onprem supervisorctl -s unix:///tmp/supervisor.sock status

# Export logs for support
./run.sh logs export

Share the exported logs folder with Inworld support when reporting issues.

Advanced: manual Docker run

For users who prefer to run Docker directly without run.sh:

docker run -d \
  --gpus all \
  --name inworld-tts-onprem \
  -p 8081:8081 \
  -p 9030:9030 \
  -e INWORLD_CUSTOMER_ID=<your-customer-id> \
  -v $(pwd)/service-account-key.json:/app/gcp-credentials/service-account.json:ro \
  us-central1-docker.pkg.dev/inworld-ai-registry/tts-onprem/tts-1.5-mini-h100-onprem:<version>

Ensure your key file has 644 permissions: chmod 644 service-account-key.json
The container exposes port 8081 (HTTP) and 9030 (gRPC)
Use docker ps to check container health — STATUS will show healthy when ready

# Stop and remove
docker stop inworld-tts-onprem && docker rm inworld-tts-onprem

# View logs
docker logs inworld-tts-onprem

# Check service status
docker exec inworld-tts-onprem supervisorctl -s unix:///tmp/supervisor.sock status

Benchmarking

For performance testing, see the Benchmarking guide.

FAQs

Can I use the on-premises container for production applications?

Yes. The on-premises container is designed for production workloads. To get started, contact sales@inworld.ai for access to the repository.

Why choose on-premises instead of cloud TTS?

For complete data control, low latency, and compliance with strict security or regulatory requirements.

Does any data leave my environment?

No. All text and audio processing occurs entirely within your environment.

How long does it take to deploy?

Deployment takes just a few minutes, with a brief model warm-up (~200 seconds).

Who is this best suited for?

Enterprises, governments, and regulated industries that cannot use cloud-based TTS.

What is included in the on-premises container?

In-scope:

API compatibility with Inworld public API
All built-in voices in Inworld’s Voice Library
The following model capabilities: text normalization, timestamps, and audio pre- and post-processing settings
Deployment how-to’s and latency benchmarks reproduction scripts

Out-of-scope:

Instant voice cloning features and their APIs
Voice design and its API

Get Started

Build with Realtime TTS

Best Practices

Resources

Why TTS On-Premises

Data stays in your environment

Low-latency, real-time speech

Designed for regulated industries

Enterprise-ready deployment

How it works

Performance

System requirements

Prerequisites

NVIDIA drivers

Docker

NVIDIA Container Toolkit

Google Cloud SDK

Verify prerequisites

Firewall requirements

Quick start

1. Create a GCP service account

3. Authenticate to the container registry

4. Configure

5. Start

6. Verify the deployment

7. Send a test request

List available voices

Lifecycle commands

Available images

Configuration

onprem.env

Logs

Troubleshooting

Advanced: manual Docker run

Benchmarking

FAQs

​Why TTS On-Premises

Data stays in your environment

Low-latency, real-time speech

Designed for regulated industries

Enterprise-ready deployment

​How it works

​Performance

​System requirements

​Prerequisites

​NVIDIA drivers

​Docker

​NVIDIA Container Toolkit

​Google Cloud SDK

​Verify prerequisites

​Firewall requirements

​Quick start

​1. Create a GCP service account

​2. Share the service account email with Inworld

​3. Authenticate to the container registry

​4. Configure

​5. Start

​6. Verify the deployment

​7. Send a test request

​List available voices

​Lifecycle commands

​Available images

​Configuration

​onprem.env

​Logs

​Troubleshooting

​Advanced: manual Docker run

​Benchmarking

​FAQs

Why TTS On-Premises

How it works

Performance

System requirements

Prerequisites

NVIDIA drivers

Docker

NVIDIA Container Toolkit

Google Cloud SDK

Verify prerequisites

Firewall requirements

Quick start

1. Create a GCP service account

2. Share the service account email with Inworld

3. Authenticate to the container registry

4. Configure

5. Start

6. Verify the deployment

7. Send a test request

List available voices

Lifecycle commands

Available images

Configuration

onprem.env

Logs

Troubleshooting

Advanced: manual Docker run

Benchmarking

FAQs