urtzai's picture
README updated
efa794c
|
raw
history blame
3.3 kB
metadata
license: cc0-1.0
language:
  - eu
tags:
  - wav
  - tts
  - piper
  - GTTS
  - CommonVoice
size_categories:
  - 10K<n<100K

Google Text-to-Speech Basque sentences

This dataset contains 100.000 Basque language audio files created automatically with Google Text-to-Speech API and sentences obtained by Mozilla's Common Voice project with Public Domain license.

Dataset Details

Dataset Description

The Basque language (as it is a minorized language) has a lack of Open Data in order to train TTS models. This models need a massive quantity of recordings that nowadays nobody in the basque open-source community possibly can afford it. So, this dataset has come to fill the gap to at least start creating basic low quality artificial intelligence (AI) models.

The Basque Common Voice dataset has 1017596 validated sentences in total. But, in this very first try with an objective of creating a synthetic Basque Open Source TTS, we just picked up a subset of first 100.000 sentences of the original dataset.

File structure

common_voice_basque_dataset.zip
metadata.json
wav/
├─ audio_1.wav
├─ audio_2.wav
├─ audio_3.wav
├─ audio_n.wav
  • common_voice_basque_dataset.zip contains 1017596 validated basque sentences used in the Mozilla's Common Voice project.
  • metadata.json file contains audio file names with each corresponding sentence.
  • wav folder contains all audio files.

Metadata

The metadata file contains the first 100.000 sentences of the Basque Common Voice dataset aligned with it's corresponding audio file name, following the Piper model training recommendations:

audio_file_name|sentences

For example the first line looks like:

audio_1|Erloju digitala du.

Starting and ending spaces of the sentences from the Basque Common Voice dataset where striped before creating the audio. So, sentences from metadata.json may differ from original Common Voice sentences.

The number from the filename it's exactly the line number of the Common Voice dataset file line number.

Audio format

The audio file format choosen was a LINEAR16, uncompressed 16-bit signed little-endian samples (Linear PCM) with a WAV header.

This choice was made because the Piper project needs wav files in order to train TTS models.

Other Google supported audio formats: https://cloud.google.com/text-to-speech/docs/reference/rest/v1/AudioConfig#AudioEncoding

Google TTS Voice

The voice used to create these audio files is the Basque Standard B female (eu-ES-Standard-B) synthetic voice. There is another voice (eu-ES-Standart-A) but the quality is lower and sounds very metallic/robotic.

More information in Google's TTS supported voices: https://cloud.google.com/text-to-speech/docs/voices

Scripts

All the scripts used to create this dataset are in this repository. Have fun! ☺️

Licenses

The entire Common Voice sentence dataset has a Public Domain license, so this Basque audio dataset does as well.