README.md

metadata

task_categories:
  - text-classification
  - text-generation
language:
  - pt

Dataset Card for 'TigreGotico/portuguese_g2p'

Table of Contents

• Dataset Description
  • Dataset Summary
  • Supported Tasks and Leaderboards
  • Languages
• Dataset Structure
  • Data Fields
  • Regional Coverage
• Dataset Creation
  • Source Data
  • Data Processing

---

Dataset Description

Dataset Summary

TigreGotico/portuguese_g2p is a Grapheme-to-Phoneme (G2P) dataset for Portuguese, offering phonetic transcriptions for sentences across ten different regional variants.

It is derived from the portuguese_phonetic_lexicon and is designed to aid in the development of robust Speech Recognition (ASR) and Text-to-Speech (TTS) models that account for dialectal variation in Portuguese.

Supported Tasks and Leaderboards

• Automatic Speech Recognition (ASR): Training language and acoustic models that require dialect-specific pronunciation.
• Text-to-Speech (TTS): Developing TTS systems capable of synthesizing speech with accurate, region-specific phonetics.
• Phonemization (G2P): Creating or evaluating G2P models that can map Portuguese text to its phonetic representation for various dialects.

Languages

The dataset is in Portuguese (pt), covering various regional and standard/non-standard dialects across Portugal, Angola, Brazil, Mozambique, and Timor-Leste.

---

Dataset Structure

Data Fields

The dataset is provided as a TSV (Tab-Separated Values) file with three main fields:

Field	Type	Description
region	string	The dialect/region code (see below for full list).
sentence	string	The written form of the sentence.
phonemes	string	The phonetic transcription of the sentence using IPA (International Phonetic Alphabet), with space as the separator between phonemes of different words. Punctuation is preserved.

Regional Coverage

The dataset explicitly models the following ten regional variants, enabling dialect-aware training:

Region	Code	Flag
Lisbon (Standard)	lbx	🇵🇹
Lisbon (Non-Standard)	lbn	🇵🇹
Luanda	lda	🇦🇴
Rio de Janeiro (Standard)	rjx	🇧🇷
Rio de Janeiro (Non-Standard)	rjo	🇧🇷
São Paulo (Standard)	spx	🇧🇷
São Paulo (Non-Standard)	spo	🇧🇷
Maputo (Standard)	mpx	🇲🇿
Maputo (Non-Standard)	map	🇲🇿
Dili	dli	🇹🇱

---

Dataset Creation

Source Data

The dataset was generated from two primary sources:

1. Lexicon Data: The base phonetic information comes from portuguese_phonetic_lexicon, which contains phonetic and morphological information for Portuguese words collected from the Portal da Língua Portuguesa.
2. Sentence Data: A collection of sentences was used as the input for the G2P process.

Data Processing

The generation pipeline involved the following steps:

1. Part-of-Speech (POS) Tagging: Input sentences were tokenized and POS-tagged using the spacy library with the pt_core_news_lg model.
2. Lexicon Lookup: For each word in a sentence, the script attempted to find its phonetic transcription in the portuguese_phonetic_lexicon, prioritizing a match based on the word, region, and predicted POS tag.
3. Fallback Phonemization:
   • If a word was not found in the lexicon, the eSpeak phonemizer was used as a fallback.
   • eSpeak was configured to use pt-BR for Brazilian variants (rjx, rjo, spx, spo) and pt-PT for all other regions, ensuring a reasonable default for missing data.
4. Assembly: The final sentence-level transcription was created by joining the phonemes of all words (excluding punctuation) with a space separator.