SpeechCommandsDataLoader<T>

Models & Types · AiDotNet.Data.Audio.Benchmarks

Loads the Google Speech Commands v2 dataset (~65K clips, 35 words, 16kHz).

Inheritance: DataLoaderBase<T> → InputOutputDataLoaderBase<T, Tensor<T>, Tensor<T>> → SpeechCommandsDataLoader<T>

How It Works

Speech Commands v2 expects the following directory structure after extraction:

Each subdirectory is a word class. WAV files are 16-bit PCM mono at 16kHz, ~1 second each.

Streaming behaviour: the loader does NOT preload audio waveforms into memory. File paths and class indices are scanned eagerly during LoadAsync, but the actual WAV decoding (and resampling, if configured) happens per-batch in ExtractBatch. This is intentional: the full 35-class dataset would allocate ~4 GB of float32 features in memory, so eager loading was guaranteed to OOM. Use GetBatches / GetBatchesAsync for training; direct Features access on this loader is intentionally not supported (features are decoded lazily per batch). Labels is a small [N, numClasses] one-hot tensor and is materialized during LoadAsync, so it remains available.

Class scheme: the “core” 12-class subset (default) follows Warden 2018: the 10 keyword classes (yes, no, up, down, left, right, on, off, stop, go) plus _silence_ (sampled from _background_noise_/) and _unknown_ (collapses every non-core word directory). The full 35-class mode uses all word directories with no synthetic classes.

Auto-download: when AutoDownload is true (default), the loader fetches and extracts the tarball from DownloadUrl if the data directory is empty, mirroring LibriSpeechDataLoader.

Constructors

Constructor Summary
SpeechCommandsDataLoader(SpeechCommandsDataLoaderOptions options) Creates a new Speech Commands data loader.

Properties

Property Summary
IReadOnlyList<string> AllWords { get; } All 35 word classes in the full dataset.
IReadOnlyList<string> CoreWords { get; } Core 10 spoken words (the standard benchmark subset).
IReadOnlyList<string> WordList { get; } Gets the word list being used (10/35 word names plus the two synthetic labels in core mode).
int FeatureCount { get; }
int NumClasses { get; } Number of word classes (12 in core mode, 35 in full mode).
int OutputDimension { get; }
int TotalCount { get; }
string Description { get; }
string Name { get; }

Methods

Method Summary
ValueTuple<IInputOutputDataLoader<T, Tensor<T>, Tensor<T>>, IInputOutputDataLoader<T, Tensor<T>, Tensor<T>>, IInputOutputDataLoader<T, Tensor<T>, Tensor<T>>> Split(double trainRatio, double validationRatio, int? seed)

Fields

Field Summary
string BackgroundNoiseDir Subdirectory containing the long background-noise WAVs distributed with the dataset.
string DownloadUrl HTTPS URL of the Speech Commands v2 archive.
string SilenceLabel Synthetic class label for background-noise / silence in the 12-class subset.
string UnknownLabel Synthetic class label for non-keyword speech in the 12-class subset.