This repository packages AudioGen as a Truss.

AudioGen is a simple and controllable model for audio generation developed by Facebook AI Research.

First, clone this repository:

git clone https://github.com/basetenlabs/truss-examples/
cd audiogen-medium-truss

Before deployment:

1. Make sure you have a Baseten account and API key.
2. Install the latest version of Truss: pip install --upgrade truss

With audiogen-medium-truss as your working directory, you can deploy the model with:

truss push

Paste your Baseten API key if prompted.

For more information, see Truss documentation.

We found this model runs reasonably fast on A10Gs; you can configure the hardware you'd like in the config.yaml.

---
resources:
  cpu: "3"
  memory: 14Gi
  use_gpu: true
  accelerator: A10G

AudioGen takes a list of prompts and a duration in seconds. It will generate one clip per prompt and return each clip as a base64 encoded WAV file.

truss predict -d '{"prompts": ['dog barking', 'sirene of an emergency vehicle', 'footsteps in a corridor'], "duration": 8}'

import json
import base64
import os, sys

model_output = json.loads(sys.stdin.read())

for idx, clip in enumerate(model_output["data"]):
  with open(f"clip_{idx}.wav", "wb") as f:
    f.write(base64.b64decode(clip))

You can also invoke your model via a REST API

curl -X POST " https://app.baseten.co/models/YOUR_MODEL_ID/predict" \
     -H "Content-Type: application/json" \
     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
     -d '{
           "prompts": ["happy rock" "energetic EDM", "sad jazz"], "duration": 8
         }'

AudioGen comes in 1 size:

• medium: 1.5B model

which is the model in this truss.