campplus-speaker

Speaker-embedding (voiceprint) bundle for IB-Robot, packaged as a schema-v3 inference_manifest.json.

One utterance in, one 192-dimensional L2-normalized vector out, so a cosine similarity is a plain dot product. IB-Robot's voice agent uses it to decide whether an utterance heard inside a wake-word session came from the person who opened that session.

Model

3D-Speaker CAM++ (speech_campplus_sv_zh-cn_16k-common), as exported to ONNX by sherpa-onnx. Upstream model card: https://www.modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common. Apache-2.0. The ONNX graph is unmodified; this bundle adds the manifest and the adapter description. See SOURCE.md for the artifact digest and how it was obtained.

Input contract

The model consumes features, not audio: 80-bin Kaldi log-mel filterbank, 16 kHz mono, 25 ms windows every 10 ms, Povey window, pre-emphasis 0.97, snip_edges=false, then per-utterance mean subtraction (global-mean). IB-Robot computes these on the host in src/voice_asr_service/voice_asr_service/speaker_id/features.py, which is pinned against sherpa-onnx's own extractor by test_speaker_features.py.

The host also splits an utterance into fixed 200-frame (2.0 s) windows, which is what the input shape above reflects. Shorter utterances are tiled from their own frames rather than zero-padded; longer ones become consecutive windows with the last aligned to the end, and the per-window unit vectors are averaged. This is part of the declared preprocessing contract, so every deployment applies it and vectors stay comparable across them.

Tensor	Semantic	Shape	dtype
input	host.campplus.features	(1, 200, 80)	float32
output	host.campplus.embedding	(1, 192)	float32

Deployments

Name	Backend	Target
torch_cpu	onnx	ONNX Runtime, CPU, 2 intra-op threads

There is no Ascend OM deployment. Measured on CANN 8.1.RC1 / Ascend310P1 (2026-10-28): the fp32 graph does not compile (Conv2D has no fp32 AI-core implementation, EZ3002), and both fp16 and mixed_float16 overflow — for identical T=200 features, 4 of 9 clips returned 10x–320x the reference norm and the remainder agreed to only 0.95–0.98, against a decision margin of 0.60. The variable-length input is a second obstacle: an OM needs a static shape, and quantizing the length into gears coarse enough to ship costs more accuracy (worst case 0.27 agreement) than the decision margin allows.

Usage

./scripts/download_models.py --models campplus-speaker

Then point a robot config at it:

voice_asr:
  speaker_embedding:
    enabled: true
    bundle_path: models/campplus-speaker
    deployment: torch_cpu

Privacy

A voiceprint is biometric data. IB-Robot keeps vectors in process memory and in the message they are published in; nothing is written to disk or kept across sessions. Deployments should keep the ROS domain isolated, since the topic is readable by any node that joins it.