ReDimNet2 ONNX

Run the official PalabraAI ReDimNet2 speaker-verification models with ONNX Runtime. All 20 published checkpoints are available and produce L2-normalized 192-dimensional embeddings.

• ReDimNet2 ONNX
  • Installation
  • Quick Start
  • Models
  • Mel Features
  • TensorRT

Installation

For raw 16 kHz audio on CPU:

pip install "redimnet2-onnx[cpu,waveform]"

Other supported setups:

Quick Start

Pass a mono 16 kHz waveform as a NumPy array or PyTorch tensor. Audio must be between 1 and 30 seconds.

from redimnet2_onnx import load_model

model = load_model("b6-vb2+vox2_v0-lm")
embedding = model.embed(waveform)

print(embedding.shape)  # (1, 192)

Models are downloaded once from the versioned GitHub Release, cached locally, and verified with SHA-256.

To compare two voices, use cosine similarity. The embeddings are already normalized:

similarity = (embedding_a @ embedding_b.T).item()

Models

List the available checkpoints:

from redimnet2_onnx import available_models, load_model

print(available_models())
model = load_model("b6-vb2+vox2_v0-lm")

Mel Features

PyTorch is optional when mel features are produced elsewhere. The ONNX input is FP32 [1, 1, 72, T], where T must be divisible by four and represent approximately 1-30 seconds of audio.

from redimnet2_onnx import load_model

model = load_model("b6-vb2+vox2_v0-lm", providers=["CUDAExecutionProvider"])
embedding = model.embed_features(features)

TensorRT

For ONNX Runtime TensorRT EP with FP16, CUDA I/O binding, and persistent engine caches:

from redimnet2_onnx.tensorrt import load_tensorrt_model

model = load_tensorrt_model("b6-vb2+vox2_v0-lm")
embedding = model.embed(waveform_tensor)

This requires compatible CUDA and TensorRT libraries on the host. The TensorRT Python package is not required.

See the documentation for the complete API reference.