vpermilp/wespeaker-voxblink2-samresnet100

WeSpeaker VoxBlink2 SimAMResNet100 (ASP)

Speaker embedding model from the WeSpeaker project, trained on the VoxBlink2 dataset.

Model Description

• Architecture: ResNet100 with SimAM (Simple, Parameter-Free Attention Module) basic blocks and Attentive Statistics Pooling (ASP)
• Block schedule: [6, 16, 24, 3] = 49 SimAM basic blocks
• Embedding dimension: 256
• Parameters: ~50.2M
• Input: Log mel-filterbank features [batch, frames, 80] (80 mel bins, 25 ms frame length, 10 ms frame shift, 16 kHz)
• Output: 256-dimensional speaker embedding [batch, 256]

Weight Format

Weights are stored as model.safetensors with PyTorch-native key naming:

Prefix	Description
front.conv1.*, front.bn1.*	Stem (3×3 conv, no maxpool)
front.layer{1-4}.{n}.conv{1,2}.*	ResNet basic blocks (49 total)
front.layer{2-4}.0.downsample.*	1×1 channel/stride projection
pooling.attention.*	ASP attention (Conv1d 5120→128→5120)
bottleneck.*	Linear 10240→256

BatchNorm parameters include weight, bias, running_mean, running_var (folded format). The num_batches_tracked bookkeeping tensor is omitted.

Usage

import torch
from safetensors.torch import load_file

state_dict = load_file("model.safetensors")
# Load into SimAM_ResNet100_ASP(in_planes=64, embed_dim=256, acoustic_dim=80)
# model.load_state_dict(state_dict, strict=True)

# Input: [B, T, 80] log-mel fbank
# Output: [B, 256] speaker embedding

Source

• Original checkpoint: avg_model.pt from the lengyue233/wespeaker-voxblink2-samresnet100 mirror
• Model code: wespeaker/models/samresnet.py
• Project page: wenet.org.cn

Re-exported from PyTorch state dict to safetensors format.