CAM++ Speaker Recognition Model (MLX)

Converted from: iic/speech_campplus_sv_zh_en_16k-common_advanced

Model Details

Architecture: CAM++ (Context-Aware Masking++)
Framework: MLX (Apple Silicon optimized)
Input: Mel-spectrogram features (320 dimensions)
Output: Speaker embedding (192 dimensions)
Quantized: False

Usage

from huggingface_hub import snapshot_download
import mlx.core as mx
import sys

# Download model
model_path = snapshot_download("mlx-community/campp-mlx")
sys.path.append(model_path)

from model import CAMPPModel
import json

# Load model
with open(f"{model_path}/config.json") as f:
    config = json.load(f)

model = CAMPPModel(
    input_dim=config["input_dim"], 
    embedding_dim=config["embedding_dim"],
    input_channels=config.get("input_channels", 64)
)
weights = mx.load(f"{model_path}/weights.npz")
model.load_weights(weights)

# Use model
audio_features = mx.random.normal((1, 320, 200))  # Your audio features
embedding = model(audio_features)

Performance

Optimized for Apple Silicon (M1/M2/M3/M4)
Faster inference than PyTorch on Mac
Lower memory usage with MLX unified memory

Original Paper

CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking https://arxiv.org/abs/2303.00332