RACON is a comprehensive feature set derived from approximately 11 hours of diverse, real-world audio. It is designed to serve as a high-quality negative dataset for training and evaluating robust wake-word models, particularly within systems like Nanowakeword.
The primary goal of this corpus is to provide developers with a challenging and realistic collection of non-target acoustic events. By incorporating RACON into a development workflow, models can learn to be more resilient and avoid unintended activations in real-world deployment scenarios.
Key Statistics
- Total Duration: ~11 hours (of source audio)
- Format: NumPy Array (
.npy) - Size (Shape):
(481345, 16, 96) - Primary Use Case: Negative Training Data, Validation, and Benchmarking
- Acoustic Content: Multi-speaker conversations, ambient noise, music, and clean speech.
Dataset Composition
The features in RACON are generated from audio sourced from three distinct and publicly available corpora, ensuring a wide acoustic diversity:
| Source Dataset | Original Duration | Description |
|---|---|---|
| DiPCo Corpus | ~5.3 hours | The entire dataset, containing challenging multi-speaker conversations in noisy home environments. |
| Santa Barbara Corpus of Spoken American English (SBCSAE) | ~3.7 hours | Selected clips featuring natural, unscripted American English conversations. |
| MUSDB18 Music Dataset | 2.0 hours | A curated selection of music clips spanning various genres. |
| Total | ~11 hours |
Purpose and Usage
RACON is a versatile negative dataset, beneficial for both the training and evaluation phases of model development.
During Training: By including these features as part of the negative class, a model learns to build a stronger decision boundary between the target wake-word and a vast array of common real-world sounds. This enhances the model's discriminative power and ability to generalize.
During Validation & Testing: Using RACON as a validation set allows for a realistic estimation of a model's performance. It helps measure the model's robustness and calculate key metrics, such as the false activation rate in a controlled environment.
Example Manifest Configuration
Below is an example of how to integrate the RACON dataset into a feature_manifest.yaml for a Nanowakeword training pipeline, where it serves as the negative validation set (negatives_val).
feature_manifest:
targets:
t: "./trained_models/tadano_A_v176/features/positive_features_train.npy"
negatives:
AE28H_float32: "./trained_models/tadano_A_v100/features/AE29H_float32.npy" # Download from here: https://huggingface.co/datasets/arcosoph/AE29H_float32
n: "./trained_models/tadano_A_v5/features/negative_features.npy"
targets_val:
tv: "./trained_models/tadano_A_v174/features/positive_features_train.npy"
negatives_val:
# RACON is used here for robust validation
bv: "RACON_11h_v1.npy"Feature Provenance and Attribution
The features in this dataset were generated using the same methodology as the openwakeword_features dataset.
Important: This dataset is a derivative work compiled for research purposes. Users must adhere to the original licensing terms of each constituent audio dataset (DiPCo, SBCSAE, and MUSDB18).
license: Apache License 2.0