So I trained a model about a week ago and when I tested it, I found that some words were mispronounced. This didn't happen on an earlier iteration of the model, so I'm worried the audio I used for the dataset could have been poisoned in some way. Is there any reason a model might mispronounce words?
#Pronunciation issues
glass plank
crude meadow
So I trained a model about a week ago and when I tested it, I found that some wo...
If it mispronounce words, is because the audio used for the dataset was noisy or you straight up used a short dataset.
glass plank
The dataset was an hour long and I cleaned it with RX beforehand.
serene wadi
The dataset was an hour long and I cleaned it with RX beforehand.
if one hour long length, bad audio inference is the only problem
use clean audio for it
crude meadow
if one hour long length, bad audio inference is the only problem
But what if he didn't delete the noise properly?
On the dataset.
serene wadi
But what if he didn't delete the noise properly?
noise is fine, as long as its barely audible, the important is clear pronunciation of words
serene wadi
On the dataset.
if talking about the dataset if noise isn't treated well, it doesn't destroy the pronunciations, those are just noise, unless they are random bleeds, not mic noise
overlapping sfx and such
also could be model and inference audio is incompatible, use only an audio with the same language or accent from the model or dataset
feral monolithBOT
Ayo? @serene wadi level 15 !!! 