Just pointing out one more thing if there is slight movement in audio with scene it could create a whole different meaning. by shifiting the position.

Hi @dineshkumar181094 great observations!

1. Could the LLM be stopping due to token limit? That might be one of the reason, as there is nothing in the prompt that is instructing LLM to restrict / stop. Ideally it should cover the whole input (85 scenes in your case).
2. In shot based indexing, where the scene duration is very short (1-2 seconds) the output might not sound coherent, here for better precision maybe better way would be to club on certain threshold (x minutes) and generate audio for those clubbed chunks instead of char based chunks given in the tutorial.
3. In our experimentation prompt generate a synced script based on the description writes a script with sentences which are roughly the same length as the time stamp of the scene in the description.

Just pointing out one more thing if there is slight movement in audio with scene it could create a whole different meaning. by shifiting the position. - Can you please share some example of this if handy? Probably good chunking should smooth out the cases like this.