README.md

metadata

license: cc-by-4.0
language:
  - en
  - fr
pretty_name: skycmd ear questions (s6-ear)
task_categories:
  - audio-classification
tags:
  - skycmd
configs:
  - config_name: clips_mswc
    data_files: data/clips_mswc.parquet
  - config_name: clips_smart-turn
    data_files: data/clips_smart-turn.parquet
  - config_name: clips_speech-commands
    data_files: data/clips_speech-commands.parquet
  - config_name: keyword
    data_files: data/keyword.parquet
  - config_name: keyword_zs
    data_files: data/keyword_zs.parquet
  - config_name: mimi_train
    data_files: data/mimi_train.parquet
  - config_name: mimi_val
    data_files: data/mimi_val.parquet
  - config_name: turn_end
    data_files: data/turn_end.parquet

skycmd ear questions (s6-ear)

The audio questions of s6-ear (turn end, keyword spotting, zero-shot keywords), the index of every clip and the Mimi codes of its audio pretraining.

Part of skycmd, tiny decision models trained from scratch in Rust that run in the browser. Written by sky-data fetch-audio and sky-data audio-jev at commit 84c6d854. Licences of every set: DATA_LICENSES.md.

• the question configs (turn_end, keyword, keyword_zs) have the typed-decision schema of skycmd-decisions (src, t, ins, opts, state, target, split) plus lang and clip; state is the path of the clip's float16 log-mel features (.npy) relative to the question file;
• the clips_* configs index every clip: id, set, split, lang, mel (features path), frames, label, and wav for the test clips;
• the mimi_* configs give, per LibriSpeech utterance, its features path and its Mimi codes.

The audio is not in this repository (31 GB of log-mel features and 16 kHz test WAVs): it is rebuilt from the upstream sets by sky-data fetch-audio, deterministically, with the same paths.

Sources and licences

Each config is one source and keeps its source's licence (share-alike sources stay share-alike). Attribution goes to the upstream authors; the conversion is ours.

config	upstream	licence	how it is used	rows
clips_mswc	MLCommons/ml_spoken_words	cc-by-4.0	clip index	360,000
clips_smart-turn	pipecat-ai/smart-turn-data-v3-train and v3.2-test	cc-by-4.0	clip index: id, split, language, frames, label	258,552
clips_speech-commands	Speech Commands v0.02	cc-by-4.0	clip index	50,289
keyword	Speech Commands v0.02 (Warden 2018), official splits and test archive	cc-by-4.0	keyword choice, the 12-class benchmark	50,289
keyword_zs	MLCommons/ml_spoken_words (en, fr; first archive of each split)	cc-by-4.0	keyword_zs choice, 50 test words per language held out	224,287
mimi_train	openslr/librispeech_asr (clean, train.100) and shangeth/librispeech-mimi-codes	cc-by-4.0	Mimi codebook-0 ids (12.5 Hz) per utterance, the audio pretraining target	28,539
mimi_val	openslr/librispeech_asr (clean, validation) and shangeth/librispeech-mimi-codes	cc-by-4.0	same, validation	2,703
turn_end	pipecat-ai/smart-turn-data-v3-train (train, val) and pipecat-ai/smart-turn-data-v3.2-test (test)	cc-by-4.0	turn_end noul questions	258,552

Files

file	rows
data/clips_mswc.parquet	360,000
data/clips_smart-turn.parquet	258,552
data/clips_speech-commands.parquet	50,289
data/keyword.parquet	50,289
data/keyword_zs.parquet	224,287
data/mimi_train.parquet	28,539
data/mimi_val.parquet	2,703
data/turn_end.parquet	258,552

from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-ear", "clips_mswc")

skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream sets carry their own teacher soft labels, see their cards), the games' oracles or our generators.