ReDimNet2-B6 ONNX Speaker Embeddings

ReDimNet2-B6 produces local speaker embeddings for comparing clean voice samples. It does not diarize audio or assign names by itself.

Model

Property	Value
Parameters	12.3 million
Format	ONNX opset 18, Float32
Model size	48.9 MiB
Input	`audio`, `[1, 96000]` mono Float32 samples
Sample rate	16 kHz
Window	6 seconds
Output	`embedding`, `[1, 192]` L2-normalized Float32

Applications should repeat clean two-to-six-second speech to fill the input and center-crop longer samples. Do not use overlapping, mixed, or unalignable speech as identity evidence.

Performance

The export is rejected unless its embedding has cosine similarity at least 0.9999 with the pinned PyTorch checkpoint and remains unit-normalized.

Measurement	Result
Warm six-second CPU inference	245.5 ms

Latency is measured on the export host and is not a Windows hardware claim.

C++ usage

The supported native host is speech-core:

#include <speech_core/models/onnx_redimnet_speaker_embedding.h>

speech_core::OnnxReDimNetSpeakerEmbedding model(
    "ReDimNet2B6.onnx");
auto embedding = model.embed(samples.data(), samples.size(), 16000);

Files

File	Description
`ReDimNet2B6.onnx`	Fixed-shape speaker encoder
`config.json`	Graph contract, provenance, hashes, and parity
`README.md`	This model card
`LICENSE`	Upstream MIT license

Source

Converted from the official PalabraAI/ReDimNet2 B6 vb2+vox2_v0 large-margin checkpoint. The pinned source revision and checkpoint SHA-256 are recorded in config.json.

Speaker embeddings are useful for labeling; they are not biometric authentication and do not protect against voice spoofing.