[FEATURE] Make Subitles instead of just Transcript

[deadcoder0904]

Problem

Sometimes I need to watch a video with subtitles and I thought about using the transcription which finished in just a second of a 15-minute video, but it was in .txt when I wanted .srt

Proposed solution

Would love .srt file with <8 words per line (or configurable) like this:

1
00:00:00,900 --> 00:00:03,800
I'm leaning into my YouTuber aesthetic now because I'm going

2
00:00:03,800 --> 00:00:06,580
to give you a reaction video to a blog post.

You can check different formats like .srt, .vtt / settings in Turboscribe fwiw.

Alternatives considered

No response

[grohith327]

Thanks for the concrete example. FluidVoice's contribution policy asks new feature ideas to begin in an Ideas Discussion unless they already have maintainer acceptance or roadmap priority.

Please carry this proposal into a Discussion, including:

• the desired output formats (.srt, .vtt, or both);
• whether cue length should be configurable by words, characters, or duration;
• which speech model produced the transcript; and
• whether initial support limited to models that expose reliable timing data would be useful.

This needs product and technical scoping because the current shared file-transcription result/export path retains final text but not subtitle cue timings. If the Discussion is accepted or prioritized, it can be converted into a tracked implementation issue.

Closing this issue as not planned for now so Issues remain focused on accepted, actionable work.

[grohith327]

Lol, was trying something with Codex 😅. Anyways, lets move this to a discussion

[deadcoder0904]

the desired output formats (.srt, .vtt, or both);

I use .srt only. I don't even know what .vtt stands for or how I should open it but its popular. Mostly needed for watching videos are home movies, sometimes that are foreign / international with no subs.

whether cue length should be configurable by words, characters, or duration;

I'd follow TurboScribe's settings since its a popular transcription SaaS built around just that. I also have a Go script on this which uses Groq API Key + Whisper 3 to transcribe stuff fast in chunks. I think it uses some popular lib like fast-whisper or something

which speech model produced the transcript

Fluid Voice. Now I think that might be an issue. I don't know if it supports transcribing to .srt or not but it did give me .txt fast in an instant fwiw on my Mac M4.

whether initial support limited to models that expose reliable timing data would be useful.

For transcriptions, timing is very important. No manual tweaking required. I don't like word level time stamps personally, but I prefer 5 to 8 words that are single on single line or sometimes on 2-lines max. Some models do like 3-4 lines or something.

[LagaV]

Would like to have vtt-subtitle format support. AFAIK this is dependent on model, e.g. Parakeet V3 does not provide timestamps, so can't output in any meaningful way to a subtitle format.

[deadcoder0904]

Oh yeah probably. I was doing large-scale transcriptions of 400 files over 3-4 hours & had to figure this out. Think I used Nova's $300 free trial for it (idk which company but a popular one) & then now use Parakeet or Whisper. There's some libraries out there that do it as well.

I have some Go scripts that work well:

cmd/generate_subtitles_groq_whisper/
cmd/generate_subtitles_parakeet_mlx/
cmd/generate_subtitles_qwen3_asr/
cmd/generate_video_subtitles_using_deepgram/ (I guess it was Deepgram's Nova)

But that day, I thought I should check out transcription but it only gave .txt but it did so real fast so was wondering if .srt is supported or not.