First 3 words getting cut off

[iskng]

Try this sentence with any voice.

We have entered a moment in history where artificial intelligence stands poised to redefine our very conception of progress.

[lucasjinreal]

Some style actually might have issues. However some style is not. I think this might caused by kokoro model training dataset of certain speaker.

Can u take a deep look at different speakers?

[iskng]

First complete word from the sentence
"We have entered a moment in history where artificial intelligence stands poised to redefine our very conception of progress."

af: 6 history
af_bella: 6 history
af_nicole: 3 a
af_sarah: 5 in
af_sky: 4 moment
am_adam: 4 moment
am_michael: 4 moment
bf_emma: 2 entered
bf_isabella: 2 entered
bm_george: 2 entered
bm_lewis: 3 a

[lucasjinreal]

the first several words is muted or relatively small volume? Does all sentences like this

[mrorigo]

I tried a hack in ort_koko.rs, and prepended the token sequence with 3 'silent' tokens, and it seems to get rid of this problem:

    pub fn infer(
        &self,
        tokens: Vec<Vec<i64>>,
        styles: Vec<Vec<f32>>,
    ) -> Result<ArrayBase<OwnedRepr<f32>, IxDyn>, Box<dyn std::error::Error>> {

        let mut tokens = tokens;
        let mut first_entry = tokens[0].clone();
        for _ in 0..3 {
            first_entry.insert(0, 30);  // token 30 seems to be kinda silent..
        }
        tokens[0] = first_entry;

@lucasjinreal idk if this is a viable solution? wfm, but might introduce a short pause when inferring multiple sentences..

[lucasjinreal]

@mrorigo This is a solution, however, I would suggestion append outside ort, since it is for universal inference only. The tokens can be appended outside as well.

Would u consider make a PR for the feature?