IndexTTS Cloud for Runpod

Important

This repository is in early development. The current launcher is a tested prototype, not yet the promised one-click Gum and PowerShell experience. It will not create anything in Runpod unless you explicitly run the launcher with a configured account.

Launch an IndexTTS2 GPU Pod from WSL2, wait for it to become ready, and print the URL and API key that a local application can use.

The public address is provided by Runpod's HTTPS proxy:

https://<pod-id>-8000.proxy.runpod.net

The default IndexTTS API key is intentionally simple:

sk-1234notsecure

What the launcher does

launchIndex.sh calls the Runpod REST API directly. It:

1. Creates a one-GPU Pod from the project's prebuilt image.
2. Tries several compatible 24 GB-or-larger NVIDIA GPU types.
3. Exposes the container's port 8000 through Runpod's HTTPS proxy.
4. Waits while the image starts, checkpoints download, and IndexTTS loads.
5. Prints INDEXTTS_URL, INDEXTTS_API_KEY, and the Pod ID.

It does not require SSH, Cloudflare Tunnel, Tailscale, Docker Compose, or Docker inside the Pod.

One-time setup

1. Publish the container image

This repository includes a GitHub Actions workflow that builds and publishes the image to GitHub Container Registry.

Push the repository to GitHub, run Build IndexTTS Cloud image, and make the resulting GHCR package public. Its address will look like:

ghcr.io/YOUR_GITHUB_NAME/indextts_cloud:latest

Alternatively, build and push it from WSL2 with Docker:

docker build -t docker.io/YOUR_DOCKER_NAME/indextts-cloud:latest .
docker login
docker push docker.io/YOUR_DOCKER_NAME/indextts-cloud:latest

The build pins upstream IndexTTS to commit 13495845e3028f0bb6ca1462ad22aa0e76349e40, the official model snapshot to 740dcaff396282ffb241903d150ac011cd4b1ede, and uses CUDA 12.8. Model weights are downloaded to /workspace on the first Pod start, so later starts reuse them.

2. Configure WSL2

cp .env.example .env
nano .env
chmod +x launchIndex.sh

Set these two values in .env:

RUNPOD_API_KEY=your-real-runpod-api-key
INDEXTTS_IMAGE=ghcr.io/YOUR_GITHUB_NAME/indextts_cloud:latest

RUNPOD_API_KEY is the only sensitive credential. Do not commit .env.

Launch

From WSL2:

./launchIndex.sh

The shell file is only a convenience wrapper. The launcher uses Python's standard library, so this is equivalent:

python3 launchIndex.py launch

The first start can take a while because Runpod must pull the large CUDA image and download the IndexTTS checkpoints. The final output looks like:

INDEXTTS_URL=https://abc123-8000.proxy.runpod.net
INDEXTTS_API_KEY=sk-1234notsecure
RUNPOD_POD_ID=abc123

The launcher records the active Pod in .runpod-index.json to prevent accidental duplicate Pods.

Manage the Pod

./launchIndex.sh status
./launchIndex.sh stop
./launchIndex.sh start
./launchIndex.sh terminate --yes

stop releases the GPU while retaining the Pod's /workspace volume. terminate --yes permanently deletes the Pod and its attached Pod volume.

Useful launch overrides:

./launchIndex.sh launch --gpu-types "NVIDIA GeForce RTX 4090,NVIDIA GeForce RTX 3090"
./launchIndex.sh launch --cloud-type COMMUNITY
./launchIndex.sh launch --spot
./launchIndex.sh launch --no-wait

API

Interactive documentation is available at:

https://<pod-id>-8000.proxy.runpod.net/docs

Authentication accepts either header:

Authorization: Bearer sk-1234notsecure
X-API-Key: sk-1234notsecure

Upload a reusable reference voice

curl "$INDEXTTS_URL/v1/voices" \
  -H "Authorization: Bearer $INDEXTTS_API_KEY" \
  -F "voice_id=default" \
  -F "file=@voice-reference.wav"

Reference audio is fundamental to IndexTTS voice cloning. The current upstream repository does not bundle an example voice, so upload one clip after the first deployment. It persists in /workspace/indextts/voices across Pod stops and starts.

You can upload additional voices the same way with IDs such as hero or narrator.

Generate speech

curl "$INDEXTTS_URL/v1/audio/speech" \
  -H "Authorization: Bearer $INDEXTTS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"indextts-2","input":"Hello from the cloud.","voice":"default"}' \
  --output speech.wav

POST /tts is an alias for the same request. Change "voice":"default" to another uploaded voice ID when needed.

Health and discovery

GET /healthz       process health; no key required
GET /readyz        model readiness; no key required
GET /v1/info       capabilities and saved voices
GET /v1/models     OpenAI-style model listing
GET /v1/voices     saved voice listing
POST /v1/voices    upload or replace a reference voice
POST /v1/audio/speech
POST /tts

WAV, MP3, Opus, and FLAC output are supported through the response_format field. Normal game dialogue should fit Runpod's HTTP proxy timeout; very long narration should be split into shorter requests.

Local verification

The API tests use a fake TTS engine and do not download models or require a GPU:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements-dev.txt
pytest -q

Actual IndexTTS inference, CUDA compatibility, and first-boot duration still need one real Runpod GPU test after the image is published.