Ikibazo gihuza ibibazo · Hagati Urwego

AI Ijwi Ryashushanyije Kuva Inyandiko Ibisobanuro

Gerageza uburyo igishushanyo cyijwi gitandukana nijwi rya clon, uburyo ibisobanuro-byanditse byamahugurwa byubatswe, nimpamvu amajwi yagenewe agomba kubikwa kugirango akomeze.

Ikibazo 1 Bya 8

Ni irihe tandukaniro nyamukuru riri hagati yo gushushanya amajwi no gutondeka amajwi?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni irihe tandukaniro nyamukuru riri hagati yo gushushanya amajwi no gutondeka amajwi?
    • Voice design needs more sample audio than cloning
    • Voice design creates a new voice from a text description, while cloning reproduces a specific speaker from reference audio
    • Cloning works only in English
    • Voice design can only make robotic-sounding voices
  2. Ni ukubera iki amakuru yo guhugura ari ikibazo gikomeye kubijyanye no kwerekana amajwi?
    • Speech audio is illegal to store
    • Models need speech paired with descriptions, and hand labeling large amounts is expensive
    • Descriptions must be written in code
    • Voice models cannot learn from more than one speaker
  3. Nigute 2024 Stability AI na Edinburgh uburyo bwo gukora ibisobanuro mubipimo?
    • By hiring voice actors to describe themselves
    • By automatically measuring attributes like pitch and speaking rate and having a language model turn them into sentences
    • By copying descriptions from audiobook covers
    • By using only synthetic speech
  4. Amabwiriza abwira ijwi ryateganijwe kugirango yumvikane neza cyane ahindura iki?
    • Voice identity
    • Delivery
    • The underlying codec
    • The speaker's accent training data
  5. Kuki ibisekuru bibiri biva mubisobanuro bimwe bishobora gutanga amajwi atandukanye?
    • The model samples one of many voices consistent with the text
    • The description is ignored after the first use
    • Audio codecs randomly change language
    • The model always adds background music
  6. In production, how do you keep a designed voice consistent across an audiobook?
    • Rewrite the description for every chapter
    • Save the chosen speaker embedding, latent or reference clip and fix the seed
    • Lower the audio sample rate
    • Switch to a different model each chapter
  7. In Parler-TTS, how does the description influence generation?
    • It is converted into a reference audio clip
    • It is encoded by a frozen T5-family encoder and used through cross-attention
    • It sets only the volume of the output
    • It is appended to the end of the waveform
  8. Which statement about designed voices is accurate?
    • A designed voice can never resemble a real person
    • Designed voices need no disclosure because no one was cloned
    • A designed voice can resemble a real person by chance, and many services block prompts naming real people
    • Designed voices can only speak one sentence