Better speech recognition

[KurohanaJuri]

Our speech recognition have some flaws and can be improved, here I will list some idea that we have (But still need more research):

• Use a phonetic algorithm
  • Pros :
    • This algo translate a word into phonetic. With this we didn't base on the spelling but how the word is pronounced.
    • "Copy-paste" an existing algorithm, this means that it's quite fast to implement this solution
    • We Can improve this with the Levenshtein distance to find the closest match
  • Cons :
    • This algo is developed for English word, It is hard to internationalized it
• Fine-tuned the existing model
  • Pros :
    • More specific to our problem and context
    • Faster than training a new model from scratch
  • Cons:
    • Time -> need a lot more research to how to fine-tune and use it inside our application
    • Dataset : I don't thing that we have a dataset that correspond our need. We need to create it ourselves
    • We don't know if we can achieve a better result. This kind of model use the context (Previous, next words) to estimate the probability of the current word. Given that our input is very short (4 words), we can say the context is almost inexistent.
    • Internationalization

We don't have THE solution for the moment.