---
language:
- gl
license: other
size_categories:
- 100K<n<1M
task_categories:
- audio-classification
- automatic-speech-recognition
task_ids:
- speaker-identification
- intent-classification
- named-entity-recognition
- audio-intent-classification
pretty_name: FalAI
tags:
- house-commands
- e-health
- question-answering
license_name: license
extra_gated_prompt: 'Terms of Access:

  By accessing and using the dataset provided herein, you agree to the terms and conditions
  of this License Agreement.

  1. Data Overview: The dataset includes anonymized voice samples labeled with personal
  information such as age, gender, place of origin, and accent.

  2. Usage Restrictions: Access to the dataset is permitted solely for research purposes
  and the development of solutions in the Galician language, including commercial
  use. However, dissemination of the voice recordings in an open-access manner or
  their public exposure is strictly prohibited. Researchers must cite the paper specified
  in the dataset description when using the dataset.


  This License Agreement is a legal contract between you and the dataset provider.
  By accessing and using the dataset, you acknowledge and agree to abide by the terms
  and conditions set forth herein.'
extra_gated_fields:
  Name: text
  Email: text
  Organization: text
  Dataset Use: text
  I accept the terms of access: checkbox
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: sentence
    dtype: string
  - name: sentence_id
    dtype: string
  - name: validation_label
    dtype: string
  - name: user_id
    dtype: string
  - name: age
    dtype: string
  - name: accent
    dtype: string
  - name: gender
    dtype: string
  - name: locality
    dtype: string
  - name: province
    dtype: string
  - name: domain
    dtype: string
  - name: intent
    dtype: string
  - name: slots
    dtype: string
  splits:
  - name: validated
    num_bytes: 52435527652.4
    num_examples: 238800
  - name: changed
    num_bytes: 3359756206.624
    num_examples: 10846
  - name: fewer_words
    num_bytes: 1132601165.912
    num_examples: 4168
  - name: hesitation
    num_bytes: 670777920.382
    num_examples: 1673
  - name: more_words
    num_bytes: 583872259.436
    num_examples: 1661
  - name: noisy
    num_bytes: 2576750236.138
    num_examples: 8041
  - name: other
    num_bytes: 115190762.0
    num_examples: 393
  download_size: 70846709348
  dataset_size: 60874476202.892006
configs:
- config_name: default
  data_files:
  - split: validated
    path: data/validated-*
  - split: changed
    path: data/changed-*
  - split: fewer_words
    path: data/fewer_words-*
  - split: hesitation
    path: data/hesitation-*
  - split: more_words
    path: data/more_words-*
  - split: noisy
    path: data/noisy-*
  - split: other
    path: data/other-*
---

# Dataset Card for FalAI Dataset


## Dataset Description


- **Curated, funded and shared by:** GTM - University of Vigo / Balidea S.L.
- **Paper:** [FalAI: A Dataset for End-to-end Spoken Language Understanding in a Low-Resource Scenario]()
- **License:** [Other](https://huggingface.co/datasets/GTM-UVigo/FalAI/blob/main/license)

### Dataset Summary

The FalAI dataset consists of a total of 265,603 audio files (wav) with associated annotations in the form of metadata.
The FalAI dataset is designed for SLU (Spoken Language Understanding) and is the **largest publicly released dataset, in any language, for the task of SLU**.
Metadata is available for each recording, including the reference phrase, validation label, user id, demographics such as age, accent, gender, locality and province, and target task metadata in the form of domain, intent and slots.

The initial text corpus consists of a total of 3,500 sentences of varying domain.
The sentences consist of standard commands or requests to virtual assistants for contexts such as house commands, e-health or e-government.
The FalAI data collection was conducted during a specific campaign in the first quarter of 2023.
Citizens were invited to participate by recording themselves reading thirty sentences using a specially [designed tool](https://falai.balidea.com/).

A careful validation of the 250 hours of audio in the dataset has been carried out, including automatic, semi-automatic and manual validations.
This has resulted in obtaining some novel splits for noisy audio, audio with hesitations, or audio with transcripts other than the reference sentence but preserving the structured information in the form of domain, intent, and slots.
These novel splits will allow the establishment of a benchmark to test SLU models in borderline scenarios (noisy environments) or more realistic scenarios (audios with hesitation), as well as to test the adaptability and generalisation of the models (modifications of the original sentence with the same structured information, dialectal variations of the language, accent variations, etc.).

### Supported Tasks

- `spoken-language-understanding` and `audio-intent-classification`: The original task for which the dataset was designed and created is to train End-to-End models for Spoken Language Understanding (E2E SLU). In these models the structured information is extracted directly from speech, in the form of domains, intents and slots.
- `audio-classification`: The dataset can be used to train audio classification models that can classify between different domains, different intents, accents, ages or other metadata present in the dataset.
- `automatic-speech-recognition`: The dataset can be used to train a model for Automatic Speech Recognition (ASR). The model is presented with an audio file and asked to transcribe the audio file to written text. The most common evaluation metric is the word error rate (WER). 
- `speaker-identification`: The dataset can also be used to train speaker identification models, using the user metadata present in the dataset.
- `intent-classification` and `named-entity-recognition`: The dataset can be used to train a model for intent or domain classification, for named-entity recognition, from audio or text.

## How to use

The `datasets` library allows you to load and pre-process your dataset in pure Python, at scale. The dataset can be downloaded and prepared in one call to your local drive by using the `load_dataset` function. 

For example, to download the validated split:
```python
from datasets import load_dataset

falai = load_dataset("GTM-UVigo/FalAI", split="validated")
```

Using the datasets library, you can also stream the dataset on-the-fly by adding a `streaming=True` argument to the `load_dataset` function call. Loading a dataset in streaming mode loads individual samples of the dataset at a time, rather than downloading the entire dataset to disk.
```python
from datasets import load_dataset

falai = load_dataset("GTM-UVigo/FalAI", split="validated", streaming=True)

print(next(iter(falai)))
```

### Local

```python
from datasets import load_dataset
from torch.utils.data.sampler import BatchSampler, RandomSampler

falai = load_dataset("GTM-UVigo/FalAI", split="validated")
batch_sampler = BatchSampler(RandomSampler(falai), batch_size=32, drop_last=False)
dataloader = DataLoader(falai, batch_sampler=batch_sampler)
```

### Streaming

```python
from datasets import load_dataset
from torch.utils.data import DataLoader

falai = load_dataset("GTM-UVigo/FalAI", split="validated")
dataloader = DataLoader(falai, batch_size=32)
```

## Dataset Structure

### Data Instances

Each data instance has the following structure, where there is data that may or may not have been shared.

```python
{
  'audio': {'path': '642c071cbd67bfe023ee279b.wav',
    'array': array([0.        , 0.        , 0.        , ..., 0.00360107, 0.00418091,
           0.00430298]),
    'sampling_rate': 44100
  },
  'sentence': 'Podes avisarme que o vinte de abril é o aniversario do veciño?',
  'sentence_id': '0bedec1a-3ef6-4184-9a76-110207c416c4',
  'validation_label': 'validated',
  'user_id': '0af70a7c-21cb-4d22-89a2-759a676bd59d',
  'age': 'range-40_49',
  'accent': 'not-shared',
  'gender': 'female',
  'locality': '32050',
  'province': '32',
  'domain': 'lists',
  'intent': 'add_to_calendar',
  'slots': "{
    'event': 'aniversario do veciño',
    'date': 'vinte de abril'
  }"
}
```

### Data Fields

* `audio` (`dict`): A dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: `dataset[0]["audio"]` the audio file is automatically decoded and resampled to `dataset.features["audio"].sampling_rate`. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the `"audio"` column, *i.e.* `dataset[0]["audio"]` should **always** be preferred over `dataset["audio"][0]`.
* `sentence` (`string`): The sentence the user was prompted to speak.
* `sentence_id` (`string`): Unique sentence identifier.
* `validation_label` (`string`): Label assigned during validation. See data splits for more information.
* `user_id` (`string`): Unique user identifier.
* `age` (`string`): The age range of the speaker (they may not have shared this information).
* `accent` (`string`): Accent of the speaker between *atlantic*, *neo-speaker*, *normative*, *eastern* or *not-shared*.
* `gender` (`string`): The gender of the speaker between *male*, *female*, *non-binary* or *not-shared*.
* `locality` (`string`): The locality of the speaker (they may not have shared this information).
* `province` (`string`): The province of the speaker (they may not have shared this information).
* `domain` (`string`): The specific topic or subject matter that the sentence pertains to.
* `intent` (`string`): The purpose or goal expressed by the speaker, which the system aims to infer or classify.
* `slots` (`string`): The specific pieces of information that the system aims to extract or identify. May be empty.

### Data Splits


The FalAI dataset introduces novel splits generated during manual validation using additional labels. The labels used during manual validation are:

* **Validated**: Utterance that exactly matches the reference sentence.
* **Changed**: Utterance in which some word(s) have been changed or pronounced differently, but the semantic information in terms of intent and slots is retained.
* **More words**: Utterance in which words are added, but the structured information is retained.
* **Fewer words**: Utterance in which words are omitted but the structured information is retained.
* **Hesitation**: Utterance in which there is a hesitation in pronunciation, whether or not the reference sentence has changed, but in which the semantic information is retained.
* **Noisy**: Noisy recording, background noise or audio problems.
* **Other**: Intelligible recordings or recordings in which the utterance has no relation to the reference sentence.

|             | Files       | Hours      |
| ----------- | ----------- | ---------- |
| Validated   | 238,800     | 222.32     |
| Changed     | 10,846      | 11.28      |
| More words  | 1,661       | 1.82       |
| Fewer words | 4,168       | 3.62       |
| Hesitation  | 1,673       | 2.15       |
| Noisy       | 8,041       | 8.47       |
| Other       | 393         | 0.36       |
| **Total**   | **265,582** | **250.02** |

## Dataset Creation

### Curation Rationale

<!-- Motivation for the creation of this dataset. -->

Galician, which belongs to the Romance language family, is a co-official language, along with Spanish, in the autonomous region of Galicia, located in northwestern Spain, and has approximately 1.9 million speakers.
Moreover, Galician is a language that has linguistic variations depending on the geographical area and that coexists in a situation of bilingualism and code-switching.

Despite its rich cultural tradition and the fact that it is the official language in public institutions, the digital presence of Galician is scarce.
For the specific case of end-to-end spoken language understanding, there is no existing dataset of speech recordings for Galician, so FalAI is the first public dataset with these characteristics.
If we analyse the available speech resources (although they are not suitable for E2E SLU, they can be used with traditional ASR + NLU architectures), the resources available in Galician are also scarce.

### Source Data

<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->

The text corpus of the FalAI dataset was designed and created in collaboration with linguists in an attempt to capture all the variants of the language, to correct and revise any errors that may have been introduced, and to establish criteria for its creation.
The FalAI dataset consists of a total of 3,500 sentences in 14 domains, with 85 intents, 34 slot types and 1,848 different values.
The domains in the dataset include classic voice command phrases for smart home speakers, but also domains related to e-health, transport booking or questions about administrative processes.

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

The FalAI data collection was conducted during a specific campaign in the first quarter of 2023. Citizens were invited to participate by recording themselves reading thirty sentences using a specially [designed tool](https://falai.balidea.com/).
The number of sentences per recording session was determined to strike a balance between a minimum number of recordings and participant motivation, typically taking 2-4 minutes per round.
Participants were encouraged to use their natural Galician, acknowledging the language's rich variations beyond the standard form.

To participate in FalAI, users were required to provide information regarding their age, gender, place of origin, and accent.
There is no limit to the number of participations, and each unique contribution counts as an additional participation, as we do not retain personal data or associated identifiers of participants.
The data collection campaign has been an unprecedented success for the language, surpassing all expectations.
It has resulted in an astounding 6 times more hours of audio data compared to the main existing speech datasets in Galician.
This campaign has achieved remarkable representation, with the participation of 99% of the municipalities in Galicia, with recordings from 311 of the 313 Galician municipalities.
The dataset includes more than 25,000 recordings per province and also captures the rich diversity of Galician accents, with more than 25,000 recordings for each of the main accent variations.
In particular, the campaign succeeded in involving a significant number of participants over the age of 60, a demographic typically underrepresented in such initiatives, with more than 15,000 recordings from this age group.


#### Annotation process

<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->

The validation process for the FalAI dataset was carefully designed due to the unique challenges posed by Galician, including the sensitivity of meaning to single character changes and the complexity of sentences containing numbers, municipality names or acronyms.
Unlike English ASRs, Galician ASRs do not have comparable confidence scores.
As a result, we opted for a semi-automatic validation strategy.

The validation strategy of the FalAI dataset included five different phases:

1. Manual validation: In the first phase, 12,750 recordings (approximately 5% of the dataset) underwent manual validation, creating a representative corpus with fully supervised validation.
2. ASR fine-tuning: In the second phase, the XLS-R model was fine-tuned using Galician speech data from Common Voice and OpenSLR . About 30% of the dataset (75,000 recordings) with 0% word error rate (WER) were automatically validated.
3. Language model boost: In the third phase, a 4-gram model trained on the dataset text corpus was used as the language model to boost recognition. In this phase, an additional 10% of the original dataset was validated with zero WER.
4. ASR model refinement: In the fourth phase, the XLS-R model was further fine-tuned using the recordings validated in the first phase. This new ASR model, reinforced by the 4-gram model, validated an additional 30% of the original dataset, with 75% of the dataset validated by the end of this phase.
5. Manual validation (final phase): The final phase manually validated recordings that hadn't been validated in previous phases. 

#### Who are the annotators?

<!-- This section describes the people or systems who created the annotations. -->

The dataset was validated manually with the collaboration of Balidea's R&D&I team.


## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

Please cite this paper when referencing the FalAI dataset:

```
@inproceedings{pineiro-martin-etal-2024-falai-dataset,
    title = "{F}al{AI}: A Dataset for End-to-end Spoken Language Understanding in a Low-Resource Scenario",
    author = "Pineiro-Martin, Andres  and
      Garcia-Mateo, Carmen  and
      Docio-Fernandez, Laura  and
      Lopez-Perez, Maria del Carmen  and
      Gandarela-Rodriguez, Jose",
    editor = "Calzolari, Nicoletta  and
      Kan, Min-Yen  and
      Hoste, Veronique  and
      Lenci, Alessandro  and
      Sakti, Sakriani  and
      Xue, Nianwen",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italy",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.624",
    pages = "7107--7116",
    abstract = "End-to-end (E2E) Spoken Language Understanding (SLU) systems infer structured information directly from the speech signal using a single model. Due to the success of virtual assistants and the increasing demand for speech interfaces, these architectures are being actively researched for their potential to improve system performance by exploiting acoustic information and avoiding the cascading errors of traditional architectures. However, these systems require large amounts of specific, well-labelled speech data for training, which is expensive to obtain even in English, where the number of public audio datasets for SLU is limited. In this paper, we release the FalAI dataset, the largest public SLU dataset in terms of hours (250 hours), recordings (260,000) and participants (over 10,000), which is also the first SLU dataset in Galician and the first to be obtained in a low-resource scenario. Furthermore, we present new measures of complexity for the text corpora, the strategies followed for the design, collection and validation of the dataset, and we define splits for noisy audio, hesitant audio and audio where the sentence has changed but the structured information is preserved. These novel splits provide a unique resource for testing SLU systems in challenging, real-world scenarios.",
}

```