Is this page helpful?

Known Limitations#

This page lists known limitations and behaviors for the Studio Voice NIM.

Audio#

  • Studio Voice expects a single mono speech stream per session. Multi-channel inputs must be downmixed on the sender, or one NIM must be deployed per speaker channel.

  • Only the 48 kHz Low-Latency model variant (studio_voice_low_latency-48k) is shipped with the Studio Voice NIM. Other Studio Voice variants (Quality 48 kHz, Quality 16 kHz) are part of the gRPC NIM and are not currently exposed in this release.

NMOS#

In NMOS mode, the receiver must be connected to the Studio Voice NIM before connecting the sender via the NMOS Connection Manager UI. Connecting the sender first can result in stale audio buffering on cold start.

For NMOS connections in the Connection Manager web UI, refer to NMOS Workflow.

Kubernetes Operator#

  • The Operator chart deploys only the controller. Sender and receiver sample workloads are provided by the End-to-End Demo Chart, not by the operator alone.

  • Deleting the NvidiaStudioVoiceMediaFunction custom resource definition while instances exist, or without uninstalling operands first, can strand resources. Follow the uninstall order—custom resources first, then the Helm release—as described in Uninstall.

  • Per-CR image override is not supported in the current release. All NvidiaStudioVoiceMediaFunction instances managed by a single operator deployment share the operator-level mediaFunction.image and mediaFunction.imagePullSecrets set at the chart level; the CR does not expose spec.image or spec.imagePullSecrets. The chart’s mediaFunction.image is a single fully-qualified image reference. To run multiple custom resources side-by-side against different NIM image tags, run multiple operator releases (each with its own mediaFunction.image); a single release can be retargeted with helm upgrade --reuse-values --set mediaFunction.image=<image-reference>, which restarts the controller and recreates managed pods. Per-CR image override might be evaluated for a future release.

Model Cache Persistence#

The model cache persistent volume claim requires a storage class that supports ReadWriteOnce access mode. When using a shared filesystem, ensure that only one pod writes to the cache concurrently to avoid corruption.

See Also#

  • Support Matrix — Supported GPUs, software versions, and resource requirements.

  • Advanced Usage — Troubleshooting common deployment issues.