Exporting

--out writes the transcript to a file; --format picks the shape:

--format	Output
plain (default)	header comments + timestamped lines
vtt	WebVTT, with speaker cues
html	standalone HTML: numbered paragraphs, timestamps, speakers
json	{"language", "text", "segments":[{start,end,speaker,text}]}
srt	SubRip subtitles: numbered cues, HH:MM:SS,mmm
ass	Advanced SubStation Alpha: one script with a Default style

transcribe jfk.wav --model whisper-tiny --out jfk.vtt --format vtt

WEBVTT
NOTE source: /audio/jfk.wav
NOTE audio 00:00:11 model whisper/whisper-tiny lang en diarize on window 300s generated 2026-01-01 09:00:00

00:00:00.000 --> 00:00:10.500
<v Speaker 0>And so my fellow Americans ask not what your country can do for you, ask what you can do for your country.</v>

NOTE done 2026-01-01 09:00:00

Real output, with the input path and timestamps replaced by placeholders. The NOTE source line records the absolute path of the input file, as does the # source header of the plain format.

JSON

--format json writes the language, the full text and the segments, with times in seconds:

{
  "language": "en",
  "text": "And so my fellow Americans ask not what your country can do for you, ask what you can do for your country.",
  "segments": [
    {"start":0,"end":10.5,"speaker":0,"text":"And so my fellow Americans ask not what your country can do for you, ask what you can do for your country."}
  ]
}

Line endings in the exported file are \n on every platform — TranscriptWriter does not go through TextWriter.WriteLine, which would emit \r\n on Windows — so a transcript exported on Windows diffs cleanly against one exported on Linux.

Subtitles: SRT and ASS

--format srt and --format ass write the same segments as cues, for a video player or a subtitle editor:

1
00:00:00,000 --> 00:00:10,500
And so my fellow Americans ask not what your country can do for you, ask what you can do for your country.

The ASS form is one script with a Default style (Arial 48, bottom-centred), so it opens in VLC, mpv or Subtitle Edit without a style to pick first. Both formats use the model’s own segment boundaries: nothing is re-timed, split or merged to make cues look shorter. A model that returns one long segment therefore produces one long cue.

Speaker labels are included only when attribution was requested (--no-diarize removes them); with them, SRT prefixes Speaker N: and ASS puts the same prefix in the dialogue text. Both are plain renderings of SpeakerSegments/Segments, not a second diarization pass.

The renderer is src/Shared/SubtitleWriter.cs, compiled into the CLI, the server and the desktop app, so the bytes do not depend on which front end exported them. The desktop app offers the same two formats in its export menu, and uses the speaker names typed in the tab when there are any.

HTML

--format html writes one self-contained document: no external CSS, no images, no JavaScript. It opens in a browser and imports into a word processor, which is what the literal line numbers in the first column are for — they are table cells, not a CSS counter, so Word keeps them.

<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="generator" content="TranscribeCppSharp">
<title>jfk.wav — transcript</title>
...
</head>
<body>
<h1>jfk.wav</h1>
<p class="meta">
source: /audio/jfk.wav<br>
audio: 00:11 · model whisper/whisper-tiny · language en · diarize on · window 300s · segments 2<br>
generated: 2026-01-01 09:00:00
</p>
<table class="transcript">
<tr><td class="ln">1</td><td class="text"><span class="ts">[00:00:00]</span> <span class="speaker">Speaker 0:</span> And so my fellow Americans ask not what your country can do for you.</td></tr>
<tr><td class="ln">2</td><td class="text"><span class="ts">[00:00:10]</span> <span class="speaker">Speaker 2:</span> Ask what you can do for your country.</td></tr>
</table>
</body>
</html>

Elements removed for brevity are marked ...; the stylesheet is a handful of rules for text, the number column and the timestamp. One table row is one segment, numbered from 1.

Timestamps ([hh:mm:ss]) are inserted on the first segment, on every speaker change, and at most once per interval — 60 s by default, so a long uninterrupted turn still carries a position every minute while the text stays readable. The interval is a setting in the desktop app’s export options; 0 leaves only the first segment and the speaker changes stamped. The CLI writes the 60 s default.

Pauses are opt-in, in the desktop app only: a gap of at least the chosen threshold (1 s by default) between two segments becomes its own unnumbered row, with one dot per whole second of silence, capped at ten. The dots are a marker of the silence between recognized segments; silence inside a segment is not visible to this export, and the transcript itself is never rewritten to include them.

Speaker names are opt-in as well: the app’s Speakers tab lists one field per speaker id, and a name typed there replaces Speaker N in the HTML and WebVTT exports. A blank field keeps Speaker N. The transcript text is untouched — a name is a label on the export, not a correction of what the model said.

The document records the model, the language, whether diarization was requested, the window and the segment count; it records nothing about accuracy, because this project publishes no accuracy figure.

Text that looks like markup is treated as text. The transcript is whatever the audio said, so it can contain <, &, { or a backtick — and the same is true of a speaker name typed in the app and of the audio file’s own name. The HTML and WebVTT exports escape those characters (<, &), the ASS export escapes its {...} override blocks, and the Markdown note puts the source, model and language inside a code fence long enough to survive a backtick in the value. SRT, plain and JSON keep the text, but on one line per segment: a line break inside a cue would end it, and let the rest of the sentence pose as a new cue or as a header. The same text printed to the console is stripped of control characters (ESC, CR, DEL, the C1 range and the invisible bidi overrides), because a terminal reads those as instructions rather than as text. None of this changes what was transcribed; it changes only how a consumer is allowed to read it.

What it is not. This is a static report, not an editable transcript. The desktop app has no editor: there is no audio-follows-cursor playback and no search-and-replace over the transcript. The line numbers exist so a reader can cite a position, not so the file can be edited in place.

html, vtt, srt and ass apply to a transcription (ASR) model only; the CLI refuses them for a model that serves another role and points at plain or json instead.

Speaker attribution in the output

With a diarization model, the attribution appears per format as Speaker N:, <v Speaker N> or "speaker":N. Which models can do it, and what you get on one that cannot, is on Long audio and speaker diarization.