Creating a dataset for automatic phonetic transcription in brazilian portuguese

Abstract

Automatic Phonetic Transcription (APT) is the technology that automates the process of converting speech into phonetic transcriptions. It is crucial for improving the accuracy of Automatic Speech Recognition (ASR) systems. Deep learning frameworks, such as wav2vec 2.0, have shown remarkable performance in learning phonetic features from data. However, they require phonetically transcribed speech corpora, which are scarse in languages such as Brazilian Portuguese (PT-BR). The primary objective of this research is to establish a systematic approach for generating a dataset with automatic phonetic transcriptions for PT-BR from available ASR corpora. By leveraging Grapheme-to-Phoneme (G2P) conversion tools, the aim is to streamline the transcription process and enhance the training of APT models. We researched PT-BR speech corpora suitable for training APT models, ultimately selecting the CORAA ASR corpus. Additionally, we evaluated five G2P converters for PT-BR, standardizing the transcriptions according to a reference phoneme chart. FalaBrasil’s G2P achieved the lowest discordance rate among the selected G2P tools, leading to its selection for transcribing the CORAA ASR corpus used in fine-tuning. The fine-tuning on 10 hours of audio yielded a 15.87% PER (Phoneme Error Rate) on the test set. In addition, the model presented high average confidence scores per phoneme, as well as little confusion between phonemes. It was then shared on the Hugging Face repository, contributing to ASR research in PT-BR.