CORAA ASR: a large corpus of spontaneous and prepared speech manually validated for speech recognition in Brazilian Portuguese

Abstract

Automatic Speech recognition (ASR) is a complex and challenging task. In recent years, there have been significant advances in the area. In particular, for the Brazilian Portuguese (BP) language, there were around 376 h publicly available for the ASR task until the second half of 2020. With the release of new datasets in early 2021, this number increased to 574 h. The existing resources, however, are composed of audios containing only read and prepared speech. There is a lack of datasets including spontaneous speech, which are essential in several ASR applications. This paper presents CORAA (Corpus of Annotated Audios) ASR with 290 h, a publicly available dataset for ASR in BP containing validated pairs of audio-transcription. CORAA ASR also contains European Portuguese audios (4.6 h). We also present a public ASR model based on Wav2Vec 2.0 XLSR-53, fine-tuned over CORAA ASR. Our model achieved a Word Error Rate (WER) of 24.18% on CORAA ASR test set and 20.08% on Common Voice test set. When measuring the Character Error Rate (CER), we obtained 11.02% and 6.34% for CORAA ASR and Common Voice, respectively. CORAA ASR corpora were assembled to both improve ASR models in BP with phenomena from spontaneous speech and motivate young researchers to start their studies on ASR for Portuguese. All the corpora are publicly available at https://github.com/nilc-nlp/CORAA under the CC BY-NC-ND 4.0 license.

Identificação de silabação em áudios de leitura de crianças em anos iniciais

“…O dataset utilizado no refinamento, chamado CORAA (Corpus of Annotated Audios) v1, é composto por 290,77 horas de áudios em português brasileiro. O refinamento do modelo, realizado por [Junior et al 2021], se deu com a GPU NVIDIA TESLA V100 32GB, usando um tamanho de batch de 8 e acumulação de gradiente sobre 24 etapas. Para auxiliar na transcrição, construiu-se um modelo de língua, que é um arquivo auxiliar que permite o ajuste de pequenas variações na transcrição segundo um dicionário de possíveis combinações de palavras.…”

AI-based approach for transcribing and classifying unstructured emergency call data: A methodological proposal

“…Our methods tested three open-source ASR deep learning fine-tuning pretrained models: Wav2Vec 2.0, HuBert, and WavLM [13][14][15]. Notably, these models are trained on large publicly available English language datasets and there is a lack of similar datasets in Brazilian Portuguese [16]. These existing models were fine-tuned to work with Brazilian Portuguese emergency calls by using the original hyperparameters from the pretrained models.…”

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop

“…Existing resources for Portuguese are composed of audios containing only read and prepared speeches, and there is a lack of datasets that include spontaneous speeches, essential in different applications. An exception is a new dataset in Portuguese designated as CORAA (Junior et al, 2021) that is composed of five different of European and Brazilian Portuguese conversations. They tried to bridge the gap of lack of spontaneity and formal speech by having only real conversations.…”