Troubleshooting#

NVCF hosting limitations#

  • After 30 seconds of inactivity the grpc channel will close.

  • You might need to re-upload your audio file when creating a grpc channel:

Currently RAM is being used to store audio in each instance. Multiple instances of the A2F-3D Authoring Microservice are hosted on NVCF. Which means that when establishing a new connection you might not be connected to the instance you uploaded your audio file to. So you will then need to upload your audio file again. The Microservice will throw an error if the audio file is not found.

Startup failures#

If you have used previous versions of the A2F-3D Authoring microservice with docker, the volumes might be cached and contain old models and SDK versions. An error message might look similar with:

$ [A2E SDK] [ERROR] Cannot Initialize from Json file: /opt/nvidia/a2f_pipeline/a2e_data/nets/a2e_v1.4.2/a2e_ms_config.json

Follow the Docker documentation to delete the old cache volumes. If you used our Container Deployment guide, run the following commands to delete the volumes and containers:

$ docker compose down -v
$ docker container prune -f

Known issues#

  • Non-verbal human sounds (E.g. “hmmmm…”) and non-human audio do not translate well into facial expressions, resulting in random lip motions. This is an area identified for future improvement.

  • At service startup the first N requests (N = selected batch size, default N=10), will have a high latency. E.g.: 160ms on a 4090 instead of 30ms.

  • Currently there is an audio file size limit of 4MB. This will be fixed with a proto file update converting the audio upload RPC to use streaming.

  • We recommend users upload audio at a 16kHz sample rate. If not, the service will perform resampling; however, the audio must not be too short to allow for successful resampling.

  • There are warnings appearing when calling generate_code_from_protos.py. These warnings are harmless.

On this page