mediainbox/redimnet2-b6-onnx

ReDimNet2 B6 ONNX

ONNX export of PalabraAI ReDimNet2 B6 using the lm vb2+vox2_v0 checkpoint. It produces normalized 192-dimensional speaker embeddings.

NOTE: Checkout redimnet2-onnx to use this model easily out-of-the-box.

Model input/output

input:  features   FP32 [1, 1, 72, T]
output: embedding  FP32 [1, 192]

The input is a 72-bin mel spectrogram, not raw audio. Audio must be mono 16 kHz. T is dynamic, divisible by four, and supports approximately 1 to 30 seconds.

Usage

from redimnet2_production import ReDimNet2

model = ReDimNet2.from_pretrained()
embedding = model.embed(waveform)

SHA-256: cdc69a3498f9d954cab263971df8faf3b97eae15197120dffb89119efb4fa21f