Allow more granularity of text length currently selected for reading

[maxstarkenburg]

Currently the extensions seems to select multiple sentences at once to read, so if I ever need to use the back button to repeat something I didn't catch the first time I listened, it'll typically go back to the beginning of the paragraph, which is often too far back (then I have to wait for it to read a long amount of text again, when I usually just want the previous sentence repeated). It would be nice to either go sentence-by-sentence, or have an option to choose one's preferred granularity (maybe including other options like word count).

As mentioned in the other issue I just opened, I use the extension on Fennec (a Firefox fork), on Android (14). I also use my local TTS (a pre-built .apk from k2-fsa/sherpa-onnx, specifically US-amy-medium, and configured to run at 2x).

However, I suspect my Voice selection might also be a factor here, since if I switch to the Google Translate voice, it seems to select things sentence-by-sentence. But I'd like to avoid sending data to Google (and I find the ReadAloud Generic Voice more robotic than I prefer, and I don't think its granularity is better anyway).

(As a side-note, it certainly seems ironic that the Google and Generic options seem much faster when I use the back/forward buttons, given that they use network, whereas my local TTS will often spend a lot of time buffering or loading (spinning icon) if I skip around too much. Perhaps I have something configured poorly, or perhaps the sherpa-onnx APKs are simply sub-optimal in that regard?)

Thanks!

[ken107]

Technically we chunk by paragraph for networking efficiency. Sending a network request for every sentence increases number of cloud API requests by several times and could increase usage cost.

Although in this case, you are using local TTS and so the network efficiency doesnt apply. Unfortunately however we dont have a way to distinguish between local vs remote voice.

The sherpa-onnx voice is performing AI inferencing locally on your computer. It's a CPU intensive operation, that's why its sluggish compared to Google voices, which are synthesized on their server farm.