quo-labs/sam-wake-word

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:

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

Usage

from datasets import load_dataset

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

License

MIT