Models

ReDimNet2 ONNX includes all 20 published ReDimNet2 checkpoints.

Model identifiers follow:

<architecture>-<dataset>-<training>

For example:

b6-vb2+vox2_v0-lm

Architecture

Architectures range from b0 through b6. Larger variants increase model capacity and compute.

Training variants

• lm - large-margin fine-tuned weights; recommended upstream for inference.
• ptn - pretrained weights before large-margin fine-tuning.

Released families

Architecture	Training data	Variants
B0-B6	VoxCeleb2	ptn, lm
B6	VoxBlink2 + VoxCeleb2	ptn, lm
B3, B6	VoxBlink2 + VoxCeleb2 + CN-Celeb2	ptn, lm

The examples in these docs use b6-vb2+vox2_v0-lm. Select the checkpoint and verification threshold using validation data representative of your application.

Discover models

Use the installed package as the source of truth for exact model identifiers:

from redimnet2_onnx import available_models

for name in available_models():
    print(name)

Inspect metadata for one checkpoint with:

from redimnet2_onnx.models import get_model_spec

spec = get_model_spec("b6-vb2+vox2_v0-lm")

print(spec.model_name)
print(spec.dataset)
print(spec.train_type)
print(spec.source_url)

For upstream benchmark results and the original PyTorch checkpoints, see the official ReDimNet2 repository.

Validate release artifacts

Maintainers can compare every exported model with the pinned PyTorch source and TensorRT FP16:

uv run python scripts/validate_tensorrt.py \
  artifacts/models-v1 \
  /path/to/voice-benchmark \
  --max-speakers 0 \
  --max-recordings-per-speaker 0 \
  --resume

The audio directory must contain one folder per speaker with at least two 16 kHz recordings. Zero limits use all qualifying data. Set TENSORRT_VERSION when it cannot be detected from the installed packages or host environment.