SpeechCommandsDataLoaderOptions

Options & Configuration · AiDotNet.Data.Audio.Benchmarks

Configuration options for the Google Speech Commands v2 data loader.

How It Works

Google Speech Commands v2 contains ~65,000 one-second audio clips of 35 spoken words recorded by thousands of different speakers at 16kHz. The “core” 12-class subset (yes, no, up, down, left, right, on, off, stop, go, silence, unknown) is the standard benchmark, with CNN baselines achieving ~95% accuracy.

Dataset: https://download.tensorflow.org/data/speech_commands_v0.02.tar.gz (2.3GB) Paper: https://arxiv.org/abs/1804.03209 License: Creative Commons BY 4.0

Constructors

Constructor Summary
SpeechCommandsDataLoaderOptions()

Properties

Property Summary
DatasetSplit Split { get; set; } Dataset split to load.
bool AutoDownload { get; set; } Automatically download and extract the dataset if it isn’t present at DataPath.
bool UseCoreSubset { get; set; } Use the core 12-class subset (yes, no, up, down, left, right, on, off, stop, go, _silence_, _unknown_) instead of all 35 classes.
int SampleRate { get; set; } Target audio sample rate in Hz.
int SilenceSampleCount { get; set; } Train-equivalent number of _silence_ clips to synthesize from the background-noise directory when the core 12-class subset is in use.
int TargetLength { get; set; } Target audio length in samples at SampleRate.
int? MaxSamplesPerClass { get; set; } Optional maximum number of samples to load per class.
string DataPath { get; set; } Root data path.

Methods

Method Summary
void Validate() Validates that all option values are within acceptable ranges.

Fields

Field Summary
int NativeSampleRate Native sample rate of the Speech Commands corpus (16kHz).