Ikibazo gihuza ibibazo · Hagati Urwego

Ikibazo cya Wav2Vec 2.0

Gerageza Wav2Vec 2.0 yiyobowe na resept yo kwitegura, intego itandukanye, n'imbibi za label yayo yo hasi ya LibriSpeech ibisubizo.

Inzira zijyanye no kuyoboraWav2vec 2 0
Ikibazo 1 Bya 8

Ni ikihe kibazo cy'ibanze mu kumenyekanisha imvugo Wav2Vec 2.0 yagenewe gukemura?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni ikihe kibazo cy'ibanze mu kumenyekanisha imvugo Wav2Vec 2.0 yagenewe gukemura?
    • Ubuke bwamajwi yanditseho (transcript)
    • Amadosiye y amajwi ari manini cyane kubika
    • Microphone zihenze cyane
    • Inyandiko iragoye kwerekana
  2. Ni ubuhe buryo bwo kwiga Wav2Vec 2.0 ikoresha mugihe cyo kwitegura?
    • Ibyiciro bya fonema itaziguye hamwe na labels yuzuye
    • Igihombo gitandukanya ibice byihishe
    • Kwiyubaka
    • Kwiga gushimangira ibitekerezo byabantu
  3. Nibihe bice byambere bitunganya imiterere mbisi muri Wav2Vec 2.0?
    • LSTM isubirwamo
    • Ikwirakwizwa
    • Inzego nyinshi za CNN ziranga kodegisi
    • Gukuramo intoki MFCC
  4. In a reported LibriSpeech experiment, what small labeled-data setup demonstrated low-label potential?
    • 10 minutes unlabeled with no pretraining
    • 53,000 hours labeled with no pretraining
    • One hour labeled with no unlabeled data
    • 10 minutes labeled after pretraining on 53,000 hours of unlabeled speech
  5. Ni uruhe ruhare rwigitabo cyize muri Wav2Vec 2.0?
    • Ihitamo amajwi ahoraho mubice byamagambo bitagira ingano
    • Irabika ijambo ryibanga ryabakoresha
    • Ihagarika inyandiko yanyuma
    • Ihitamo inyungu za mikoro
  6. Ni ubuhe bwoko bunini bw'indimi nyinshi bwubatswe ku buryo bwa Wav2Vec 2.0?
    • GPT-2
    • XLS-R
    • Stable Diffusion
    • ResNet-50
  7. What does the Transformer contribute during Wav2Vec 2.0 pretraining?
    • Converts transcripts into audio
    • Clusters final text into language labels
    • Contextualizes latent speech features around masked spans
    • Removes the feature encoder
  8. What is typically done after self-supervised pretraining to create an ASR model?
    • Replace the audio encoder with an image classifier
    • Train only on unlabeled text
    • Discard the model and use k-means alone
    • Fine-tune with transcribed speech