ww datasets

_posts/2025-11-01-ww_datasets.md

@@ -0,0 +1,251 @@
+---
+title: "Building Better Wake Words: Why Data Matters and How You Can Get Started"
+excerpt: "Wake word systems live and die by the data they’re trained on. A model trained on a few speakers in quiet rooms might work fine in the lab but fail the moment it hears a new accent, microphone, or background noise.  High-quality, diverse datasets are the foundation of any robust wake word system. They capture different voices, devices, and environments, helping models generalize and reducing false triggers or missed activations."
+coverImage: "/assets/blog/ww_datasets/thumb.png"
+date: "2025-11-01T00:00:00.000Z"
+author:
+  name: JarbasAl
+  picture: "https://avatars.githubusercontent.com/u/33701864"
+ogImage:
+  url: "/assets/blog/ww_datasets/thumb.png"
+---
+
+
+## Building Better Wake Words: Why Data Matters and How You Can Get Started
+
+### What Are Wake Words?
+
+Wake words — short trigger phrases like *“Hey Computer,” “Alexa,”* or *“Hey Mycroft”* — are the key that wakes up your voice assistant. They signal when the system should start listening and processing commands.
+
+A good wake word detector should respond quickly and accurately to the right phrase — and stay silent the rest of the time. That simple goal hides a complex challenge: real-world speech is unpredictable, noisy, and incredibly diverse.
+
+---
+
+### Why Quality Data Is Essential
+
+Wake word systems live and die by the data they’re trained on. A model trained on a few speakers in quiet rooms might work fine in the lab but fail the moment it hears a new accent, microphone, or background noise.
+
+High-quality, diverse datasets are the foundation of any robust wake word system. They capture different voices, devices, and environments, helping models generalize and reducing false triggers or missed activations.
+
+But collecting this kind of data is difficult — and responsibly gathering diverse speech samples at scale is even harder. That’s why open, shareable datasets are so important.
+
+---
+
+### Common Challenges in Wake Word Research
+
+Developers and researchers often run into the same issues when building wake word systems:
+
+* **Data scarcity** – Few open datasets exist for wake word detection.
+* **Imbalance** – Far more “non-wake” samples than “wake” ones.
+* **Noise sensitivity** – Background sounds and reverb can break performance.
+* **Speaker diversity** – Models often fail to generalize to new voices or accents.
+
+OpenVoiceOS has been organizing a growing collection of open datasets — both wake word–specific and general audio corpora — to make it easier for anyone to build, train, and benchmark detection systems.
+
+---
+
+## 🗣️ Wake Word Datasets
+
+### Synthetic Data
+
+Generated specifically for controlled experiments, these synthetic and augmented datasets include speech created using TTS and voice cloning. These datasets are a work in progress, expect more data to be added over time.
+
+👉 [Synthetic Wake Word Collection](https://huggingface.co/collections/TigreGotico/synthetic-wakeword-datasets)
+
+---
+
+### Mycroft / OpenVoiceOS Community Data
+
+Real-world samples collected by the Mycroft and OpenVoiceOS communities. These recordings include multiple speakers, accents, and recording setups.
+
+👉 [OVOS Community Wake Words Dataset](https://huggingface.co/datasets/OpenVoiceOS/ovos-community-wakewords-dataset)
+
+| Wake Word        | Samples | Notes                            |
+| ---------------- | ------- | -------------------------------- |
+| `amelia`         | 223     |                                  |
+| `athena`         | 482     |                                  |
+| `computer`       | 470     | used to train Precise-Lite model |
+| `cristopher`     | 3       |                                  |
+| `ey_ordenador`   | 48      | Spanish                          |
+| `hey_chatterbox` | 116     | used to train Precise-Lite model |
+| `hey_computer`   | 14      |                                  |
+| `hey_firefox`    | 3       |                                  |
+| `hey_floyd`      | 96      |                                  |
+| `hey_k9`         | 46      |                                  |
+| `hey_kit`        | 5       |                                  |
+| `hey_mike`       | 2       |                                  |
+| `hey_moxie`      | 4       |                                  |
+| `hey_savant`     | 115     |                                  |
+| `hey_scout`      | 34      |                                  |
+| `hey_ziggy`      | 26      |                                  |
+
+---
+
+### Qualcomm Snapdragon Keyword Dataset
+
+A balanced, multi-speaker dataset featuring four English keywords recorded by 50 speakers.
+
+👉 [Qualcomm Keyword Speech Dataset](https://www.qualcomm.com/developer/software/keyword-speech-dataset)
+
+| Wake Word        | Samples | Notes        |
+| ---------------- | ------- | ------------ |
+| `hey_snapdragon` | 1,112   | 50 speakers  |
+| `hi_galaxy`      | 934     | 200 speakers |
+| `hi_lumina`      | 1,112   | 50 speakers  |
+| `hey_android`    | 1,112   | 50 speakers  |
+
+
+---
+
+### Google Speech Commands
+
+👉 [HuggingFace](https://huggingface.co/datasets/google/speech_commands)
+👉 [Kaggle Mirror](https://www.kaggle.com/datasets/neehakurelli/google-speech-commands)
+
+A widely used dataset for keyword spotting — also useful for building wake word or “not-wake” word datasets.
+
+| Wake Word | Samples | Notes                            |
+| --------- | ------- | -------------------------------- |
+| `marvin`  | 2,100   | used to train Precise-Lite model |
+| `sheila`  | 2,022   | used to train Precise-Lite model |
+| `stop`    | 3,872   |                                  |
+| `yes`     | 4,044   |                                  |
+| `no`      | 3,941   |                                  |
+
+---
+
+### Pico Benchmark
+
+Designed to compare wake word detection engines. Over 300 recordings per wake word from more than 50 speakers.
+
+👉 [GitHub Repository](https://github.com/Picovoice/wake-word-benchmark)
+👉 [HuggingFace Mirror](https://huggingface.co/datasets/domdomegg/picovoice-wake-word-benchmark)
+
+| Wake Word      | Samples |
+| -------------- | ------- |
+| `alexa`        | 329     |
+| `computer`     | 411     |
+| `jarvis`       | 384     |
+| `smart_mirror` | 369     |
+| `snowboy`      | 401     |
+| `view_glass`   | 399     |
+
+---
+
+### Alexa Dataset (Kaggle)
+
+A small dataset of Alexa wake word samples.
+
+👉 [Kaggle – Alexa Wake Word Dataset](https://www.kaggle.com/datasets/aanhari/alexa-dataset)
+
+| Wake Word | Samples |
+| --------- | ------- |
+| `alexa`   | 738     |
+
+---
+
+### NAR Dataset
+
+Recordings made with the Nao robot in real domestic environments — includes speech and background sounds with natural noise and reverb.
+
+👉 [NAR Dataset on Hugging Face](https://huggingface.co/datasets/TigreGotico/NAR)
+👉 [Official Page](https://team.inria.fr/perception/nard/)
+
+| Wake Word | Samples |
+| --------- | ------- |
+| `stop`    | 20      |
+| `yes`     | 20      |
+| `no`      | 20      |
+
+---
+
+### Multilingual Spoken Words Corpus
+
+Massive dataset of spoken words across 50 languages — ideal for multilingual research and cross-lingual training.
+
+👉 [MLCommons Multilingual Spoken Words](https://mlcommons.org/datasets/multilingual-spoken-words/)
+👉 [Hugging Face Mirror](https://huggingface.co/datasets/MLCommons/ml_spoken_words)
+
+---
+
+## 🔊 Auxiliary “Not Wake Word” and Noise Datasets
+
+These datasets aren’t wake words themselves — but they’re incredibly useful as *negative samples* or background noise when training detection systems.
+
+---
+
+### [DipCo – Dinner Party Corpus, Interspeech 2020](https://zenodo.org/records/8122551)
+
+👉 [Hugging Face Mirror](https://huggingface.co/datasets/huckiyang/DiPCo)
+
+Simulated dinner-party recordings with multiple speakers, overlapping speech, and background music — perfect for training models to handle real-world noise.
+
+---
+
+### [FMA (3 Seconds)](https://huggingface.co/datasets/TigreGotico/FMA_3secs)
+
+3-second clips from the Free Music Archive — ideal for background and false trigger testing.
+
+---
+
+### [MIT Environmental Impulse Response Dataset](https://huggingface.co/datasets/davidscripka/MIT_environmental_impulse_responses)
+
+271 audio files of real-world acoustic impulse responses — great for simulating room reverberation and microphone characteristics.
+
+---
+
+### [Public Domain Sounds (3 Seconds)](https://huggingface.co/datasets/TigreGotico/public_domain_sounds_3secs)
+
+635 copyright-free environmental recordings, chopped into 3-second clips — a versatile source of background sounds.
+
+---
+
+### [Freiburg 106](https://huggingface.co/datasets/TigreGotico/building_106_kitchen_3secs)
+
+Environmental recordings from a kitchen environment, split into short clips.
+
+---
+
+### [ESC-50: Environmental Sound Classification](https://huggingface.co/datasets/TigreGotico/ESC-50)
+
+2,000 labeled 5-second environmental recordings across 50 sound classes — a go-to dataset for environmental noise classification and augmentation.
+
+---
+
+### [AudioSet](https://huggingface.co/datasets/agkphysics/AudioSet)
+
+10-second clips from YouTube, labeled across hundreds of sound categories — a powerful general-purpose dataset for sound event modeling.
+
+---
+
+## Why Share All This?
+
+Open datasets are the backbone of reproducible wake word research. By sharing and cataloging these resources, we hope to make it easier for anyone — from academic researchers to hobbyists — to train, compare, and improve wake word detection systems.
+
+These collections combine real, synthetic, and environmental data that together form a strong foundation for open-source wake word development.
+
+---
+
+## What’s Next
+
+OpenVoiceOS is preparing a new round of **wake word benchmarks** that will put existing models and engines to the test across these datasets — from quiet labs to noisy kitchens.
+
+While many community projects (like Precise) have carried open wake word detection forward, they’re still **hard to train**, **data-hungry**, and **don’t generalize well** to new voices. Our upcoming work aims to change that — with better evaluation standards, more data diversity, and open comparisons.
+
+Stay tuned — benchmarks are coming soon. 👀
+
+---
+
+## Help Us Build Voice for Everyone
+
+OpenVoiceOS is more than software, it’s a mission. If you believe voice assistants should be open, inclusive, and user-controlled, here’s how you can help:
+
+- **💸 Donate**: Help us fund development, infrastructure, and legal protection.
+- **📣 Contribute Open Data**: Share voice samples and transcriptions under open licenses.
+- **🌍 Translate**: Help make OVOS accessible in every language.
+
+We're not building this for profit. We're building it for people. With your support, we can keep voice tech transparent, private, and community-owned.
+
+👉 [Support the project here](https://www.openvoiceos.org/contribution)
+
+