vggish_params.py

# MFCC Spectrogram conversion code from VGGish, Google Inc.
# https://github.com/tensorflow/models/tree/master/research/audioset

NUM_FRAMES = 96  # Frames in input mel-spectrogram patch.
NUM_BANDS = 64  # Frequency bands in input mel-spectrogram patch.
EMBEDDING_SIZE = 128  # Size of embedding layer.

SAMPLE_RATE = 16000
STFT_WINDOW_LENGTH_SECONDS = 0.025
STFT_HOP_LENGTH_SECONDS = 0.010
NUM_MEL_BINS = NUM_BANDS
MEL_MIN_HZ = 125
MEL_MAX_HZ = 7500
LOG_OFFSET = 0.01  # Offset used for stabilized log of input mel-spectrogram.
EXAMPLE_WINDOW_SECONDS = 0.96  # Each example contains 96 10ms frames
EXAMPLE_HOP_SECONDS = 0.96     # with zero overlap.

PCA_EIGEN_VECTORS_NAME = 'pca_eigen_vectors'
PCA_MEANS_NAME = 'pca_means'
QUANTIZE_MIN_VAL = -2.0
QUANTIZE_MAX_VAL = +2.0

INIT_STDDEV = 0.01  # Standard deviation used to initialize weights.
LEARNING_RATE = 1e-4  # Learning rate for the Adam optimizer.
ADAM_EPSILON = 1e-8  # Epsilon for the Adam optimizer.

INPUT_OP_NAME = 'vggish/input_features'
INPUT_TENSOR_NAME = INPUT_OP_NAME + ':0'
OUTPUT_OP_NAME = 'vggish/embedding'
OUTPUT_TENSOR_NAME = OUTPUT_OP_NAME + ':0'
AUDIO_EMBEDDING_FEATURE_NAME = 'audio_embedding'