--- tags: - kws, - edge - impulse - synthetic - voice --- # Hey Android Wake Word Synthetic Speech Dataset This dataset contains synthetic and augmented audio samples for training and testing a small wake-word / keyword-spotting model for the phrase: ```text Hey Android ``` The dataset was created as a bootstrap dataset for embedded audio classification experiments, particularly for Edge Impulse, TensorFlow Lite, and TinyML workflows. It is intended to help developers quickly prototype an Android-style wake-word recognizer before collecting larger real-world recordings. ## Dataset Summary The dataset contains three primary labels: | Label | Description | | ------------------ | --------------------------------------------------------------------------------------------------------------------------------- | | `hey_android` | Synthetic utterances of the target wake phrase and close variants such as “Hey Android”, “Hello Android”, and “Okay Android”. | | `unknown` | Near-miss and non-wake-word phrases such as “Hey Andrew”, “Hey Adrian”, “Hello phone”, “Open settings”, and other short commands. | | `background_noise` | Synthetic background noise clips used to help the model distinguish speech from non-speech audio. | The audio clips are generated as fixed-length mono WAV files suitable for keyword spotting and small embedded audio classifiers. ## Intended Use This dataset is intended for: * wake-word detection experiments * keyword spotting * TinyML audio classification * Edge Impulse audio projects * Android voice-trigger prototypes * embedded ML tutorials * speech dataset bootstrapping * model deployment testing on microcontrollers, phones, or small Linux edge devices Recommended starting classes: ```text hey_android unknown background_noise ``` Accent, speaker, voice, and locale variation should be treated as variation **inside** the class labels, not as separate labels. For example, Irish English, American English, British English, Dutch-accented pronunciation, and other synthetic voice differences should all remain part of the same `hey_android` class when the spoken phrase is the target wake phrase. ## Not Intended For This dataset is not intended to be a production-grade benchmark or final production training set. It should not be used as the only training data for a deployed wake-word product. Synthetic speech is useful for bootstrapping, but it does not fully represent: * real microphones * far-field speech * spontaneous human timing * household acoustics * car cabin audio * phone microphone compression * overlapping speakers * TV/radio background * real accent distributions * real false-trigger conditions For production use, add real recordings from the target device and expected deployment environments. ## Data Format The dataset is organised as WAV files split into training and test folders: ```text audio/ train/ hey_android..wav unknown..wav background_noise..wav test/ hey_android..wav unknown..wav background_noise..wav ``` Additional metadata files may include: ```text hf_metadata.csv metadata.csv selected_voices.csv dataset_summary.json hf_dataset_summary.json ``` `hf_metadata.csv` contains a simple file index: | Column | Description | | -------------- | --------------------------------------------------- | | `audio` | Relative path to the WAV file | | `label` | Class label | | `split` | `train` or `test` | | `filename` | WAV filename | | `source_file` | Source file used for augmentation, where applicable | | `augmentation` | Description of augmentation applied | ## Audio Specification The generated dataset is designed around the following audio settings: | Property | Value | | --------------------- | -------------------------------------- | | Format | WAV | | Channels | Mono | | Sample rate | 16 kHz | | Clip length | 2 seconds | | Primary task | Keyword spotting / wake-word detection | | Primary target phrase | `Hey Android` | ## Dataset Creation The dataset was created using a Google Cloud Text-to-Speech generation workflow and local audio augmentation. The generation workflow included: 1. Selecting locale-specific text-to-speech voices. 2. Generating target wake-word phrases and unknown / near-miss phrases. 3. Exporting clips as mono 16 kHz WAV files. 4. Applying local augmentation. 5. Creating train/test splits. 6. Writing Edge Impulse-compatible label-prefixed filenames. 7. Preparing Hugging Face metadata and dataset card files. The dataset generation process included voices/locales such as: * Irish English: `en-IE` * American English: `en-US` * British English: `en-GB` * Australian English: `en-AU` * Dutch: `nl-NL` * Belgian Dutch: `nl-BE` Availability depends on the text-to-speech provider and selected voice family. Some Google Cloud TTS voice families, especially Chirp / Chirp3 voices, may have stricter quota limits or unsupported audio parameters such as pitch. The generation workflow therefore supports: * excluding Chirp / Chirp3 voices * disabling pitch augmentation * adding request delays * retrying quota/rate-limit errors with backoff * using local augmentation to expand the dataset without additional TTS calls ## Augmentation Local augmentation may include: * gain variation * small time shifts * additive synthetic noise * simple echo / room-like effects * synthetic background noise generation This is intended to increase robustness during early-stage prototyping, but it is not a substitute for real-world data collection. ## Loading the Dataset Basic file-based loading: ```python import pandas as pd df = pd.read_csv("hf_metadata.csv") print(df.head()) ``` Example using Hugging Face Datasets: ```python from datasets import load_dataset, Audio dataset = load_dataset("eoinedge/hey-android") if "audio" in dataset["train"].column_names: dataset = dataset.cast_column("audio", Audio(sampling_rate=16000)) print(dataset) print(dataset["train"][0]) ``` If the dataset viewer does not infer the audio column automatically, use `hf_metadata.csv` to load files manually from the repository. ## Edge Impulse Usage The WAV filenames are compatible with Edge Impulse label-prefix conventions: ```text hey_android..wav unknown..wav background_noise..wav ``` You can upload the files with the Edge Impulse uploader: ```bash edge-impulse-uploader --category training audio/train/*.wav edge-impulse-uploader --category testing audio/test/*.wav ``` Recommended Edge Impulse impulse: ```text Input block: Audio, 2 seconds, 16 kHz Processing block: MFCC or MFE Learning block: Classification Classes: hey_android, unknown, background_noise ``` ## Recommended Follow-Up Data Collection This dataset should be expanded with real recordings from: * phone microphones * Android devices * laptop microphones * quiet rooms * cars * kitchens * fans * outdoor background noise * TV/radio playing nearby * close and far microphone distances * real Irish, Dutch, American, British, Indian, and other speakers * near-miss phrases such as “Hey Andrew”, “Hey Adrian”, and “Hey Androids” A stronger wake-word dataset should include both positive and negative samples recorded under the expected deployment conditions. ## Limitations Synthetic TTS data can be overly clean, consistent, and well-articulated. It may not represent: * true human variation * natural hesitation * speech interruptions * children’s voices * accented spontaneous speech * far-field recordings * low-quality microphones * compression artifacts * overlapping speakers * real acoustic environments Treat this as a bootstrap dataset, not a final benchmark. ## Ethical and Licensing Notes This dataset contains synthetic speech generated for a wake-word detection experiment. It does not intentionally include real personal speech recordings. The dataset is released under CC BY 4.0. Users should verify that their use of generated synthetic speech complies with the terms of the text-to-speech provider used to create their local copy. ## Citation ```bibtex @dataset{jordan_hey_android, author = {Eoin Jordan}, title = {Hey Android Wake Word Synthetic Speech Dataset}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/datasets/eoinedge/hey-android} } ```