This repository contains textual SLURP data and some additional resources allowing to replicate the results published in our EMNLP 2020 paper.
- Textual annotations are part of this repository and can be found under
dataset/slurp/
directory - To download the corresponding acoustic data, run
scripts/download_audio.sh
script (need around 6GB of disk space). Data is kindly hosted by Zenodo.org - an open science sharing platform.
Please feel free to get in touch if you obtain any interesting results on SLURP (and they are publicly available, either as a paper or some other form like blogs, etc.). We are happy to link to them from here.
- Results for end-to-end architecture by Lore≈n Lugosch: https://lorenlugosch.github.io/posts/2020/12/slurp
- Each SLURP entry has the following structure. It contains NLU annotations and information allowing to link with audio files, as well as the corresponding metadata information.
- Note that
wer
(Word Error Rate) andent_wer
(Entity WER) fields for each recording were obtained automatically by scoring the ASR hypotheses against expected ground truths. In the paper we used all SLURP entries to train ASR systems (including ones wherewer>0
, but feel free to use this field to filter out less certain recordings. - Naming convention for audio files encodes close and distant microphones (
-headset
suffix denotes close-talk variant)
- Note that
{
"slurp_id": 4489,
"sentence": "wake me up at eight o'clock",
"sentence_annotation": "wake me up at [time : eight] o'clock",
"intent": "alarm_set",
"action": "set",
"tokens": [
{
"surface": "wake",
"id": 0,
"lemma": "wake",
"pos": "VB"
},
{
"surface": "me",
"id": 1,
"lemma": "-PRON-",
"pos": "PRP"
},
{
"surface": "up",
"id": 2,
"lemma": "up",
"pos": "RP"
},
{
"surface": "at",
"id": 3,
"lemma": "at",
"pos": "IN"
},
{
"surface": "eight",
"id": 4,
"lemma": "eight",
"pos": "CD"
},
{
"surface": "o'clock",
"id": 5,
"lemma": "o'clock",
"pos": "NN"
}
],
"scenario": "alarm",
"recordings": [
{
"file": "audio-1502099775-headset.flac",
"wer": 0,
"ent_wer": 0,
"status": "correct"
},
{
"file": "audio-1502099775.flac",
"wer": 0,
"ent_wer": 0,
"status": "correct"
},
{
"file": "audio--1504194195-headset.flac",
"wer": 0,
"ent_wer": 0,
"status": "correct"
},
{
"file": "audio--1504194195.flac",
"wer": 0,
"ent_wer": 0,
"status": "correct"
},
{
"file": "audio-1492784000-headset.flac",
"wer": 0,
"ent_wer": 0,
"status": "correct"
},
{
"file": "audio-1492784000.flac",
"wer": 0,
"ent_wer": 0,
"status": "correct"
}
],
"entities": [
{
"span": [
4
],
"type": "time"
}
]
}
- Each SLURP record (as above) has the corresponding metadata entry in
dataset/slurp/metadata.json
, providing more information on selected items.- Many fields are for our internal use to link SLURP entries to raw database (can be igonored).
- Some, however, may be of use. For example,
usrid
field encodes speaker's gender and primary language (native / non-native). The encoding follows the patternGenderLanguage-Id
, whereGender
could be eitherF
(Female),M
(Male) orUNK
(Unknown). Language is eitherE
(native English speaker) orO
(native of Other langauge than English).Id
is a uniq three-digit number, does not carry any meaningul information. As suchFE-488
is the recording uttered by a female participant whose English is a native language.
"4489": {
"nlub_id": 5027,
"recordings": {
"audio--1504194195-headset.flac": {
"ansid": 17555,
"recid": 50412,
"status": "correct",
"usrid": "FO-488"
},
"audio--1504194195.flac": {
"ansid": 17555,
"recid": 50412,
"status": "correct",
"usrid": "FO-488"
},
"audio-1492784000-headset.flac": {
"ansid": 17555,
"recid": 20567,
"status": "correct",
"usrid": "UNK-323"
},
"audio-1492784000.flac": {
"ansid": 17555,
"recid": 20567,
"status": "correct",
"usrid": "UNK-323"
},
"audio-1502099775-headset.flac": {
"ansid": 17555,
"recid": 43759,
"status": "correct",
"usrid": "FO-460"
},
"audio-1502099775.flac": {
"ansid": 17555,
"recid": 43759,
"status": "correct",
"usrid": "FO-460"
}
},
"sentence_normalised": "wake me up at eight o'clock",
"sentence_original": "wake me up at eight o'clock",
"status": "correct",
"userid_amt": "682"
}
- Audio data. Once downloaded you will find two directories
slurp_real
andslurp_synth
underaudio
directory in the root of the repo. The files are compressed with Free Lossles Audio Codec FLAC.
If this dataset proves useful, please cite our corresponding EMNLP 2020 paper:
@inproceedings{slurp,
author = {Emanuele Bastianelli and Andrea Vanzo and Pawel Swietojanski and Verena Rieser},
title={{SLURP: A Spoken Language Understanding Resource Package}},
booktitle={{Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)}},
year={2020}
}