Teste SpecAugment para reconhecimento de fala

As respostas e explicações aparecem conforme você responde ao teste.

1. Em que o SpecAugment opera?
   • O espectrograma log-mel
   • A forma de onda de áudio bruta
   • A transcrição do texto
   • Os pesos do modelo
2. Qual destas é uma das três operações do SpecAugment?
   • Inversão de tom
   • Mascaramento de frequência
   • Substituição de palavras
   • Injeção de eco
3. What role did the SpecAugment paper assign to its feature transformations?
   • They guarantee that every recognizer avoids overfitting
   • They replace the need for transcribed training data
   • They augment ASR training inputs and were evaluated for recognition performance
   • They improve every speech model by the same amount
4. O que a 'máscara de tempo' faz?
   • Exclui os carimbos de data/hora da transcrição
   • Acelera a reprodução de áudio
   • Remove o silêncio no início
   • Anula um intervalo de intervalos de tempo consecutivos
5. How are time- and frequency-mask locations chosen during SpecAugment training?
   • They are randomly sampled, so the same input may receive different masked regions on later passes
   • They are fixed permanently before the training set is collected
   • They are chosen by a language model after each recognition result
   • They are identical for every utterance and every training step
6. Em qual benchmark o SpecAugment ajudou notavelmente a alcançar resultados de última geração?
   • ImageNet
   • LibriSpeech
   • ESQUADRÃO
   • COLA
7. What does time warping do in the original SpecAugment policy?
   • It replaces speech frames with translated text
   • It masks every frequency channel permanently
   • It deforms feature positions along the time axis
   • It changes the speaker identity in the waveform
8. Which recognizer architecture did the original SpecAugment paper use in its experiments?
   • A region-based image detector
   • A music source-separation U-Net only
   • A text-only masked language model
   • Listen, Attend and Spell (LAS)