Orpheus TTS 3B — RunPod Serverless Worker

Emotional TTS for the video pipeline via RunPod GPU cloud.

Architecture

VPS (CPU, $30/yr)                    RunPod Serverless (GPU)
┌──────────────────┐                 ┌──────────────────────┐
│ orchestrator.py   │    HTTP API     │ Orpheus TTS 3B       │
│ produce_recap.py  │ ──────────────► │ RTX 4090 (24GB)      │
│ ffmpeg            │ ◄────────────── │ ~$0.05/chapter       │
│ sound library     │    WAV audio    │ Scales to zero       │
└──────────────────┘                 └──────────────────────┘

Setup (One-Time)

1. Create RunPod Account

Go to https://www.runpod.io/ → Sign Up → Add $10 credit.

2. Build & Push Docker Image

Option A: GitHub integration (easiest)

• Push this directory to a GitHub repo
• In RunPod: Serverless → New Template → Connect GitHub repo
• RunPod builds the image automatically

Option B: Build locally (needs Docker Desktop)

cd runpod-orpheus/
docker build --platform linux/amd64 -t YOURDOCKERHUB/orpheus-tts-runpod:latest .
docker push YOURDOCKERHUB/orpheus-tts-runpod:latest

3. Create Serverless Endpoint

In RunPod Console:

1. Go to Serverless → New Endpoint
2. Select your template (from step 2)
3. GPU: RTX 4090 (24GB)
4. Workers: Min 0 / Max 1 (flex — scales to zero)
5. Idle timeout: 30s (keeps warm briefly after request)
6. Click Create

Copy the Endpoint ID (looks like abc123xyz).

4. Get API Key

RunPod Console → Settings → API Keys → Create Key.

5. Configure VPS

ssh -i ~/.ssh/yonglibrary_vps root@100.85.221.52

# Add to pipeline user's environment
cat >> /home/pipeline/.bashrc << 'EOF'
export TTS_ENGINE=orpheus
export RUNPOD_API_KEY=your_key_here
export ORPHEUS_ENDPOINT_ID=your_endpoint_id_here
EOF

# Apply patch
scp patch_orpheus_tts.py to VPS, then:
python3 /tmp/patch_orpheus_tts.py

6. Test

su - pipeline
cd /opt/video-pipeline
python3 -c "
from scripts.orpheus_client import OrpheusClient
c = OrpheusClient()
wav = c.generate('Hello, this is a test of Orpheus TTS.', voice='tara')
print(f'Got {len(wav)} bytes of audio')
with open('/tmp/test_orpheus.wav', 'wb') as f:
    f.write(wav)
print('Saved to /tmp/test_orpheus.wav')
"

Emotion Tags

Orpheus supports inline emotion tags in text:

Example: "Well, that's interesting. <laugh> I hadn't expected that."

Voices

Cost