Brazilian Portuguese Speech Recognition Using Wav2vec 2.0

Gris, Lucas Rafael Stefanel; Casanova, Edresson; Oliveira, Frederico Santos de; Soares, Anderson da Silva; Cândido, Arnaldo

doi:10.1007/978-3-030-98305-5_31

Cited by 7 publications

(4 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In Table 3 we show the performance of four ASRs trained with Brazilian Portuguese datasets, varying the amount of spontaneous and prepared speech data. [Candido Junior et al 2021] 0,221 0,132 0,534 0,436 0,742 0,503 0,241 0,164 0,228 0,104 0,457 0,281 [Ferreira and dos Reis Oliveira 2022] 0,247 0,147 0,529 0,453 0,738 0,521 0,269 0,186 0,230 0,102 0,461 0,289 [Grosman 2022] 0,265 0,168 0,523 0,484 0,754 0,564 0,265 0,180 0,201 0,105 0,451 0,304 [Stefanel Gris et al 2022] 0,333 0,186 0,656 0,516 0,859 0,600 0,319 0,203 0,321 0,150 0,588 0,365 Average 0,267 0,158 0,561 0,472 0,773 0,547 0,274 0,183 0,245 0,115 0,489 0,309 Table 3. Performance, using WER and CER, of four ASRs trained with BP data.…”

Section: Results Using Cer and Wer Metricsmentioning

confidence: 99%

“…The ASR model described in [Candido Junior et al 2021] presents the best average values of WER and CER for the five inquiries evaluated (0.393 and 0.268, respectively), while the model described in [Stefanel Gris et al 2022], which was trained only with prepared speech, has the worst performance for all the five inquiries. [Ferreira and Oliveira 2022] and [Grosman 2022] models also presented consistent results, with a small advantage for Ferreira and Oliveira's model.…”

Section: Results Using Cer and Wer Metricsmentioning

confidence: 99%

“…This section presents four ASR models evaluated for the task of transcribing NURC/SP audios: [Candido Junior et al 2021], [Ferreira and Oliveira 2022], [Grosman 2022], and[Stefanel Gris et al 2022]. These works are based on Wav2vec 2.0 and were trained for Portuguese speech recognition using the corpora presented in Section 2.2.2.…”

Section: Trained Models For Portuguesementioning

confidence: 99%

“…We also explored the role of language model in the experiments involving this model. Finally, the work of [Stefanel Gris et al 2022] is based on Wav2vec 2.0 XLSR-53 reaching 12.4% of WER evaluated against several large corpora in Portuguese, namely CETUC, MLS, Common Voice 7.0 and also some small corpora. This work was selected to contrast with the work models, since it has no spontaneous speech in its training set.…”

Section: Trained Models For Portuguesementioning

confidence: 99%

See 3 more Smart Citations

Bringing NURC/SP to Digital Life: the Role of Open-source Automatic Speech Recognition Models

Gris

Cândido

Santos

et al. 2022

Anais Do XIX Encontro Nacional De Inteligência Artificial E Computacional (ENIAC 2022)

View full text Add to dashboard Cite

The NURC Project that started in 1969 to study the cultured linguistic urban norm spoken in five Brazilian capitals, was responsible for compiling a large corpus for each capital. The digitized NURC/SP comprises 375 inquiries in 334 hours of recordings taken in São Paulo capital. Although 47 inquiries have transcripts, there was no alignment between the audio-transcription, and 328 inquiries were not transcribed. This article presents an evaluation and error analysis of three automatic speech recognition models trained with spontaneous speech in Portuguese and one model trained with prepared speech. The evaluation allowed us to choose the best model, using WER and CER metrics, in a manually aligned sample of NURC/SP, to automatically transcribe 284 hours.

show abstract

Section: Results Using Cer and Wer Metricsmentioning

confidence: 99%