quo-labs/sam-wake-word

---
language:
- en
task_categories:
- audio-classification
tags:
- wake-word
- keyword-spotting
- voice-assistant
pretty_name: SAM Wake Word Dataset
size_categories:
- 1K<n<10K
license: mit
configs:
- config_name: default
data_files:
- split: train
path: "**/*.wav"
default: true
---

# SAM Wake Word Dataset

Audio dataset for training a wake-word detection model to recognise the keyword **"Sam"**.

## Dataset Description

Each sample is a short audio clip labelled as either **positive** (contains the wake word) or **negative** (does not).

| Split | Description |
|-------|-------------|
| `positive/` | Clips of the word "Sam" spoken in varied styles, speeds, and intonations |
| `negative/` | Clips of phonetically similar or common words that are **not** "Sam" |

## Generation

All audio is synthesised using **OpenAI `gpt-4o-mini-tts`** with:

- **10 voices**: alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer
- **Variable speed**: 0.85× – 1.30×
- **Diverse speaking styles**: whispering, shouting, questioning, commanding, accented, etc.

Positive prompts include: *"Sam", "sam", "SAM", "Sam!", "Hey Sam", "Yo Sam"*

Negative prompts include phonetically close words (*ham, jam, slam, spam, dam, same, samuel, sample*) and common short words (*hello, hey, stop, play, yes, no, …*).

### Audio Format

- **Sample rate**: 16 kHz
- **Channels**: mono
- **Format**: WAV

## Manifest

A `manifest.json` file is included with metadata for each clip:

```json
{
"file": "positive/positive_00042.wav",
"label": "positive",
"text": "Sam",
"voice": "nova",
"speed": 1.12
}
```

## Usage

```python
from datasets import load_dataset

ds = load_dataset("quo-labs/sam-wake-word")
```

## License

MIT