Is this page helpful?

Pipeline Configuration#

This page covers NIM service chart Helm values and environment variables specific to the Active Speaker Detection pipeline, including tuning parameters, media format, and SMPTE ST 2110 session wiring.

Active Speaker Detection Pipeline Tuning#

Configuration	Helm Key (`pipeline.*`)	Environment Variable	Comment
Sync tolerance (minimum score to treat a face as speaking)	`syncTolerance`	`AI4M_ACTIVESPEAKERDETECTION_SYNC_TOLERANCE`	Default: `0.5986`; valid range: `(0, 1)`; invalid values fall back to default at process start.
Bounding box overlay on output video	`testFrameOverlayMode`	`AI4M_ACTIVESPEAKERDETECTION_TEST_FRAME_OVERLAY_MODE`	`true` or `false`; when enabled, draws active-speaker bounding boxes on output video.
Output frame buffer size	`outputFrameBufferSize`	`AI4M_ACTIVESPEAKERDETECTION_OUTPUT_FRAME_BUFFER_SIZE`	Default: `30` frames; adds ~`size / fps` seconds of latency to the ~3 sec model latency.
Audio silence detection	`useAudioThresholdToDetectActiveAudioStream`	`AI4M_ACTIVESPEAKERDETECTION_USE_AUDIO_THRESHOLD_TO_DETECT_ACTIVE_AUDIO_STREAM`	Default: `false`; allowed: `true` or `false`.
Audio silence threshold	`audioThresholdDb`	`AI4M_ACTIVESPEAKERDETECTION_AUDIO_THRESHOLD_DB`	Default: `-40.0` dB; used when audio silence detection is enabled.

Video Raster#

The video raster defines the fixed picture format of the incoming SMPTE ST 2110-20 stream. These Helm keys tell the NIM the raster of the incoming video so that decoding, audio sync, and inference stay consistent.

Configuration	Helm Key	Comment
Width and height	`video.width`, `video.height`	Supported: 720p, 1080p (default), 4K UHD; refer to Support Matrix.
Frame rate	`video.framerateNum`, `video.framerateDen`	Must match the upstream sender and downstream receiver.

Audio Format#

Configuration	Helm Key	Comment
PCM format	`audio.pcmFormat`	Example: `S24BE`
Sample rate	`audio.samplingRate`	Example: `48000` Hz
Channels	`audio.numChannels`	Per logical stream

SMPTE ST 2110 Input Sessions#

Configuration	Helm Key	Comment
Video session	`input.video.sessionName`, `input.video.localInterfaceName`, `input.video.hostIp`, `input.video.hostPort`, `input.video.hostNumSubnetBits`, `input.video.flushFrameCountThreshold`	Multicast IP address and port must match the upstream sender pipeline; site-specific. `flushFrameCountThreshold`: consecutive stale frames allowed before flush and resync (default: `3`).
Audio session	`input.audio.sessionName`, `input.audio.localInterfaceName`, `input.audio.hostIp`, `input.audio.hostPort`, `input.audio.hostNumSubnetBits`, `input.audio.numStreams`	Multicast IP address and port must match the upstream sender pipeline. `numStreams` must match the number of diarized audio streams.

SMPTE ST 2110 Output Sessions#

For video and ancillary outputs, multicast hostIp and hostPort must match the multicast host IP address and host port configured on the downstream receiver pipeline. (Receivers must subscribe to the same groups and ports this NIM transmits on.)

Configuration	Helm Key	Comment
Ancillary (SMPTE ST 2110-40)	`output.ancillaryData.sessionName`, `output.ancillaryData.localInterfaceName`, `output.ancillaryData.hostIp`, `output.ancillaryData.hostPort`, `output.ancillaryData.hostNumSubnetBits`	Active-speaker detection results as ancillary data. For application payload layout, refer to Ancillary Data Payload.
Video out	`output.video.sessionName`, `output.video.localInterfaceName`, `output.video.hostIp`, `output.video.hostPort`, `output.video.hostNumSubnetBits`	SMPTE ST 2110-20 program video; site-specific session name, interface, and subnet mask length.

NIM Pod Resources#

Resource	Requests	Limits
`cpu`	`12`	`12`
`hugepages-2Mi`	`8Gi`	`8Gi`
`memory`	`8Gi`	`8Gi`
`nvidia.com/gpu`	`1`	`1`

The cpu value is set to 12 cores by default, tuned for real-time ST 2110 performance. You can reduce it, but the value must be consistent with env.CPU_AFFINITY.

env.CPU_AFFINITY (default "0-7") pins the NIM process to a contiguous range of CPU cores. The chart validates at deploy time that the values of resources.requests.cpu and resources.limits.cpu are each equal to or greater than the number of cores in the affinity range.

Setting	Helm Key	Default
CPU affinity range	`env.CPU_AFFINITY`	`"0-7"` (8 cores)

The default cpu (12) is intentionally greater than CPU_AFFINITY ("0-7", 8 cores)—the pinned range covers the NIM workload minimum, and the remaining cores are available to Kubernetes. To configure the cores, keep resources.requests.cpu, resources.limits.cpu, and env.CPU_AFFINITY consistent. If cpu falls below the affinity range, the chart fails during installation with a descriptive error.

Logging#

Configuration	Helm Key	Comment
Log level	`logging.level`	`0` FATAL, `1` ERROR, `2` WARN, `3` INFO (default), `4` DEBUG, `5` VERBOSE.