How to get sentence-level timestamps?

parakeet output is word-level timestamped, but is friendly/traditional to make a subtitle for a video/audio transcription. Is possible to get sentence-level output? There is a local tool that can do it?

I was able to get word-level transcription with:

parakeet-cli transcribe --model tdt-0.6b-v3-q4_k.gguf --input in.wav --json > out.json

And with this shell script I was able to convert to WebVTT subtitle:

#!/bin/sh
jq -r '
  def format_time:
    . as $s |
    ($s / 3600 | floor) as $h |
    (($s % 3600) / 60 | floor) as $m |
    ($s % 60 | floor) as $sec |
    ($s * 1000 | floor % 1000) as $ms |
    "\(if $h < 10 then "0" else "" end)\($h):\(if $m < 10 then "0" else "" end)\($m):\(if $sec < 10 then "0" else "" end)\($sec).\(if $ms < 10 then "00" elif $ms < 100 then "0" else "" end)\($ms)";

  "WEBVTT\n\n" + ([.words[] | "\(.start | format_time) --> \(.end | format_time)\n\(.w)"] | join("\n\n")) + "\n"
' "$1"

• The output is one word per subtitle line, that is not very friendly to create a subtitle for a video or audio recording.
• Do you know a way to convert the transcription in a more traditional (sentence-level timestamps) subtitle?

Workaround

Using Word Timestamp to Subtitles online tool, I was able to get an expected formatted subtitle from JSON output (from parakeet), but I'm searching a local/offline tool, because privacy.

This should work but give me errors, and I'm not a programmer to fix that.

Start with the parakeet-cli transcribe command and its JSON output, then inspect how word timestamps are exposed. Compare the requested sentence-level subtitle behavior with the linked make_subtitles.py script and determine whether this belongs in the CLI or as a local companion tool. Done should mean a documented, privacy-preserving way to produce traditional sentence-level subtitles.