Um estudo de caso do modelo de reconhecimento de voz whisper para transcrição de conferências tedx via aprendizado fraco

Abstract

Machine learning models for speech recognition tasks are often trained with unsupervised learning because of the difficulty of collecting high-quality labeled data. The multitasking and multilingual speech recognition model called Whisper gave us the possibility to use a state-of-art model in the area of natural language processing for free. The great innovation of this model was to use weak learning, mixing a large amount of labeled data and unlabeled data, with a high degree of content diversity, demonstrating that weak learning can perform at the same level as other state-of-art models. The Whisper model’s voice recognition and transcription-to-text capabilities in Portuguese were studied by applying the Multilingual TEDx dataset, which contains more than 150 hours of high-quality audio in .flac format and entirely in Portuguese, in addition to their transcripts audios in .vtt format. We were able to observe an error rate below 1, using the Word Error Rate metric, fluctuating between 0.3 and 0.7. This rate demonstrates that the model can, when exposed to small inputs, have an error rate below that recorded in training when exposed to larger inputs.