Listen, Attend and Spell: Character Negative Log-Likelihood
Intermediate
The problem, from first principles
LAS decodes characters with an attention decoder trained by maximum likelihood. The training loss is the sum of the negative log-probabilities of the correct characters, the same sequence loss as in machine translation.
From theory to code
Implement char_nll(logp, targets), the summed negative log-likelihood of the transcript.
Constraints
- Returns a Python float.
Hints
Hint 1
Pick the log-probability of the correct character at each step, negate, and sum.
Keywords
research-paperssequence-modelsspeechattention
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.