Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

RNN-Transducer Models adanwo

Understand how RNN-Transducer models combine acoustic frames and output history for streaming sequence transcription tasks.

Jẹmọ itọsọna awọn ọnaAwọn awoṣe Oluyipada Rnn
Ibeere 1 ti 8

Eyi ti aropin ti CTC ni RNN-Transducer adirẹsi pataki?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Eyi ti aropin ti CTC ni RNN-Transducer adirẹsi pataki?
    • Iro inu-ominira ipo-ominira ti CTC, eyiti o ṣe idiwọ awọn igbẹkẹle iṣelọpọ-àmi
    • CTC ko le mu ohun ṣiṣẹ rara
    • CTC nilo awọn GPU lati ṣiṣẹ
    • CTC ṣiṣẹ fun awọn aworan nikan
  2. Kini awọn ẹya akọkọ mẹta ti RNN-Transducer?
    • Encoder, decoder, ati iyasoto
    • Encoder (nẹtiwọki transcription), nẹtiwọki asọtẹlẹ, ati nẹtiwọki apapọ
    • Tokenizer, ifibọ, ati softmax nikan
    • Awọn olori akiyesi meji ati olulana kan
  3. Kini ipa wo ni nẹtiwọọki asọtẹlẹ ṣe ninu RNN-T?
    • O compresses awọn iwe igbi fọọmu
    • O ṣe iwọn iwọn gbohungbohun
    • O ṣe bi awoṣe ede kan, ti n ṣatunṣe lori awọn ami ti o jade tẹlẹ
    • O yan iwọn ayẹwo
  4. Kini idi ti RNN-T ṣe atilẹyin ṣiṣan lairi kekere bẹ nipa ti ara?
    • O ṣe ilana gbogbo ọrọ ni ẹẹkan ṣaaju gbigbejade ohunkohun
    • O foju pa aṣẹ akoko patapata
    • O ṣiṣẹ nikan lori awọn faili ti a ti gbasilẹ tẹlẹ
    • Titete rẹ jẹ monotonic ati osi-si-ọtun, ti nlọsiwaju nipasẹ ohun pẹlu awọn ofo
  5. Ninu iyipada RNN-T, kini jijade ami-ami òfo ṣe?
    • Awọn ilọsiwaju si igbesẹ akoko ohun afetigbọ atẹle
    • Pari gbogbo transcription
    • Npa ọrọ ti tẹlẹ rẹ jẹ
    • Yipada si ede miiran
  6. Iru kooduopo wo ni o rọpo awọn LSTM ni iṣelọpọ ode oni awọn ọna ṣiṣe RNN-T?
    • Awọn igi ipinnu
    • Conformer encoders
    • Itele apo-ti-ọrọ
    • Farasin Markov Models
  7. What two histories condition the RNN-T joint network’s output distribution?
    • Only the future transcript and a speaker photo
    • Only the last acoustic frame, without output context
    • The encoder representation of acoustic input and the prediction-network representation of prior output labels
    • The optimizer state and learning rate
  8. What alignment behavior does an RNN-T allow before advancing to the next audio frame?
    • It must emit exactly one label for every frame
    • It advances time only after a word-final token
    • It generates all output before reading any audio
    • It can emit zero or more nonblank labels at the current frame, then a blank advances time