LaFresCat Multiaccent

We present LaFresCat, the first Catalan multiaccented and multispeaker dataset.

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. Commercial use is only possible through licensing by the voice artists. For further information, contact [email protected] and [email protected].

Dataset Details

Dataset Description

The audios from this dataset have been created with professional studio recordings by professional voice actors in Lafresca Creative Studio. We processed the raw audios with the following recipe:

• Trimming: Long silences from the start and the end of clips have been removed.

  • py-webrtcvad -> Python interface to the Voice Activity Detector (VAD) developed by Google for the WebRTC.

• Resampling: From 48000 Hz to 22050 Hz, which is the most common sampling rate for training TTS models.

• SOX -> SOX is a versatile command-line utility primarily used for converting audio files between different formats.

• Stereo to mono conversion: The original raw audios were provided in stereo, but for TTS training the audios neede to be in mono. Because of that, we used the libraries librosa and soundfile to make this conversion and posterior check.

• Librosa Librosa is a Python library for music and audio analysis, offering tools for feature extraction, audio manipulation, and visualization.

• Soundfile SoundFile is a library for reading and writing sound files in various formats, providing a simple interface for audio I/O operations.

In total, there are 4 different accents, with 2 speakers per accent (female and male). After trimming, accumulates a total of 3,75h (divided by speaker IDs) as follows:

• Balear

  • olga -> 23.5 min
  • quim -> 30.93 min

• Central

  • elia -> 33.14 min
  • grau -> 37,86 min

• Occidental (North-Western)

  • emma -> 28,67 min
  • pere -> 25,12 min

• Valencia

  • gina -> 22,25 min
  • lluc -> 23,58 min

Uses

The purpose of this dataset is mainly for training text-to-speech and automatic speech recognition models in Catalan dialects.

Languages

The dataset is in Catalan (ca-ES).

Dataset Structure

The dataset consists of a single split, providing audios and transcriptions:

DatasetDict({
    train: Dataset({
        features: ['audio', 'transcription', 'speaker_id', 'accent'],
        num_rows: 2858
    })
})

Each data point is structured as:

>> audio_dataset[0]

{'audio': {'path': 'lafresca_multiaccent/central/grau/grau_220.wav',
  'array': array([0., 0., 0., ..., 0., 0., 0.]),
  'sampling_rate': 22050},
 'transcription': 'Una mica més amunt, un cop passats els blocs de pisos, hi havia una altra casa de dos o tres pisos en la que plantaven floretes petites blanques.',
 'speaker_id': 'grau',
 'accent': 'central'}

Dataset Splits

• 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].

  • path (str): The path to the audio file.
  • array (array): Decoded audio array.
  • sampling_rate (int): Audio sampling rate.

• transcription (str): The sentence the user was prompted to speak.

Dataset Creation

This dataset has been created by members of the Language Technologies unit from the Life Sciences department of the Barcelona Supercomputing Center. The voices belong to professional voice actors and they've been recorded in Lafresca Creative Studio.

Source Data

The data presented in this dataset is the source data.

Data Collection and Processing

These are the technical details of the data collection and processing:

• Microphone: Austrian Audio oc818

• Preamp: Focusrite ISA Two

• Audio Interface: Antelope Orion 32+

• DAW: ProTools 2023.6.0

Processing:

• Noise Gate: C1 Gate

• Compression BF-76

• De-Esser Renaissance

• EQ Maag EQ2

• EQ FabFilter Pro-Q3

• Limiter: L1 Ultramaximizer

Here's the information about the speakers:

Dialect	Gender	County
Central	male	Barcelonès
Central	female	Barcelonès
Balear	female	Pla de Mallorca
Balear	male	Llevant
Occidental	male	Baix Ebre
Occidental	female	Baix Ebre
Valencian	female	Ribera Alta
Valencian	male	La Plana Baixa

Who are the source data producers?

The Language Technologies team from the Life Sciences department at the Barcelona Supercomputing Center developed this dataset. It features recordings by professional voice actors made at Lafresca Creative Studio.

Annotations

In order to check whether or not there were any errors in the transcriptions of the audios, we created a Label Studio space. In that space, we manually listened to subset of the dataset, and compared what we heard with the transcription. If the transcription was mistaken, we corrected it.

Personal and Sensitive Information

The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in this dataset.

Bias, Risks, and Limitations

Training a Text-to-Speech (TTS) model by fine-tuning with a Catalan speaker who speaks a particular dialect presents significant limitations. Mostly, the challenge is in capturing the full range of variability inherent in that accent. Each dialect has its own unique phonetic, intonational, and prosodic characteristics that can vary greatly even within a single linguistic region. Consequently, a TTS model trained on a narrow dialect sample will struggle to generalize across different accents and sub-dialects, leading to reduced accuracy and naturalness. Additionally, achieving a standard representation is exceedingly difficult because linguistic features can differ markedly not only between dialects but also among individual speakers within the same dialect group. These variations encompass subtle nuances in pronunciation, rhythm, and speech patterns that are challenging to standardize in a model trained on a limited dataset.

Recommendations

Citation

APA:

Funding

This work has been promoted and financed by the Generalitat de Catalunya through the [Aina project] (https://projecteaina.cat/).

Dataset Card Contact

[email protected]