ERes2NetV2 Speaker Verification Chinese 16k

This repository contains an ERes2NetV2 speaker verification model exported from the original PyTorch checkpoint to Hugging Face style safetensors format.

Model Details

• Architecture: ERes2NetV2 speaker embedding model
• Task: speaker verification and speaker embedding extraction
• Language/domain: Chinese speech
• Sample rate: 16 kHz
• Acoustic feature: 80-dimensional Kaldi fbank
• Embedding size: 192
• Weight file: model.safetensors
• Config file: config.json

The exported weights are converted from:

model-bin/speech_eres2netv2_sv_zh-cn_16k-common/pretrained_eres2netv2.ckpt

Repository Files

.
|-- config.json
|-- model.safetensors
`-- README.md

config.json keeps the model metadata used by this project and points pretrained_model to model.safetensors.

Loading

Use the project inferencer:

from campplus_sid.inferencer.inferencer import Inferencer

inferencer = Inferencer.from_pretrained(
    "model-bin/safetensor/speech_eres2netv2_sv_zh-cn_16k-common",
    device="cpu",
)

For direct weight loading:

from safetensors.torch import load_file

state_dict = load_file(
    "model-bin/safetensor/speech_eres2netv2_sv_zh-cn_16k-common/model.safetensors"
)

Input Format

The model expects normalized 80-bin fbank features with shape:

batch, frames, feature_dim

In this project, raw 16 kHz audio is converted to fbank features by the Inferencer before the model forward pass.

Reference

This architecture is associated with the 3D-Speaker speaker recognition work:

https://github.com/alibaba-damo-academy/3D-Speaker

Verification

The exported model.safetensors was validated by loading all tensors and checking the state dict against the ERes2NetV2 model definition with no missing or unexpected keys.