Seven assistant names. Seven datasets. All audio is generated entirely by the phoonnx TTS framework using the Miro and Dii voices. No human recordings. No consent forms. No privacy exposure.
Each dataset is a flat set of roughly a thousand positive clips: the wakeword spoken with varied speakers, rates, and prosody. Hard negatives and background noise ship as separate companion datasets you mix in at training time: not-wake-words-speech-en, not-wake-words-speech-pt, and ambient_noises.
Why synthetic
Real recordings take months to collect. Every speaker needs a consent form, and the result still leaves accent gaps you didn’t anticipate. Synthetic generation avoids all of that.
- Reproducible: the same generation settings and voices produce the same audio, with a full audit trail and no consent-form archaeology.
- Auditable: the generation pipeline is the documentation.
- Scalable: varying speaking rate and speaker characteristics is a parameter change, not a studio session.
For wakeword detection, what matters is acoustic distinctiveness, not naturalness. Synthetic data fits that requirement.
Use these
Train your own wakeword detector for OpenVoiceOS, Mycroft, or any open voice system. For negative-sample augmentation there are also household and public-domain background-clip datasets: building_106_kitchen_3secs, public_domain_sounds_3secs, and FMA_3secs.