AkashPrasadMishra/multilingual-keywords-900

Multilingual Keywords 900

Overview

This dataset contains isolated, single spoken-word recordings collected across multiple languages: English, Japanese, Turkish, Spanish, Czech, Portuguese, and others. Each sample is a short audio clip of exactly one spoken word — there is no fixed segment length, since word duration naturally varies by language, speaker, and phonetic length of the word itself.

• 900 unique keywords/words total, spread across the included languages
• ~200–300 audio samples per word on average (different speakers/utterances of the same word)
• Variable clip duration — no artificial trimming or padding to a fixed window; each clip is as long as the natural utterance

This makes it well suited for short-word / single-utterance classification tasks, rather than continuous-speech or sentence-level tasks.

---

Key Difference vs. Fixed-Length Datasets

Unlike windowed datasets (e.g. noise datasets sliced into fixed 1.5s chunks), clip length here is not standardized — a short word in one language may be ~300ms, while a longer word in another language may run 1–2 seconds or more. Any model consuming this data should handle variable-length input (e.g. via padding/truncation in the dataloader, or duration-aware batching) rather than assuming a fixed window.

---

Dataset Structure

multilingual-keywords-900/
├── en/
│   ├── <word_1>/
│   │   ├── sample_001.wav
│   │   ├── sample_002.wav
│   │   └── ...
│   ├── <word_2>/
│   └── ...
├── ja/
│   ├── <word_1>/
│   └── ...
├── tr/
├── es/
├── cs/
├── pt/
└── ...

(Adjust this tree to match your actual folder/naming layout if it differs — e.g. flat filenames with a metadata CSV instead of per-word subfolders.)

---

Audio Format

Property	Value
Clip length	Variable (natural word duration)
Sample rate	(fill in, e.g. 16 kHz)
Channels	(fill in, e.g. mono)
Bit depth	(fill in, e.g. 16-bit PCM)
File format	.wav

---

Dataset Statistics

Metric	Value
Languages covered	English, Japanese, Turkish, Spanish, Czech, Portuguese (+ any others — list all)
Unique keywords	900
Samples per keyword (avg)	~200–300
Total samples (approx.)	~180,000 – 270,000
Total duration	(fill in)

---

Intended Use Cases

• Keyword spotting / hotword / wake-word detection — training or evaluating models that must recognize one or a small set of trigger words.
• Language identification (LID) — since keywords span multiple languages, per-language clusters can support language classification from short audio.
• Robot / assistant "identity" or activation-word generation — building custom wake-word models for voice assistants or robotics platforms.
• Voice activity detection (VAD) — positive speech examples of short, isolated duration, useful alongside noise datasets as negatives.
• General short-word / single-utterance audio classification research.

---

Collection Notes

• Words were collected as single spoken-word utterances, not extracted from continuous speech — each file should contain one word only, without surrounding silence trimmed inconsistently. (Confirm whether leading/trailing silence was trimmed or left natural.)
• Multiple speakers contributed samples per word to introduce speaker/accent variability. (Fill in: number of unique speakers, if tracked, and whether speaker metadata is included per sample.)
• Language balance is not perfectly even — samples per word average 200–300, but this may vary by language/word.

---

Licensing

(Fill in the actual license terms for this collection — e.g. whether it's your own recordings, crowd-sourced, or derived from another corpus. If it aggregates multiple public sources, license terms may vary per language/source and should be documented per subset.)

---

Known Limitations

• No fixed clip length — downstream pipelines must handle variable-duration audio.
• Class balance across the 900 keywords and 6+ languages may not be perfectly uniform (some words/languages may have more or fewer samples).
• Speaker demographics, accents, and recording conditions (microphone, environment) are not detailed here — add this if such metadata was tracked during collection, since it affects model generalization.