Timestamps in Transcription: When and How Granular

The first question is not which granularity, but whether you need timestamps at all. A transcript for reading, sharing as notes, or feeding into a summarizer gains nothing from timestamp markers. They add visual noise and larger file size without value. Timestamps pay off in one situation: when you need to navigate back to a specific moment in the audio, or sync text to a video display.

Do You Need Timestamps at All?

Start here before picking a granularity.

If your use case lands in the "No" column, stop. Export plain text, skip the anchors, and keep the document clean.

The Four Granularities

When you do need timestamps, the right level is determined by what you will do with the result.

None

A clean paragraph transcript without any time markers. Right for reading, sharing, and text processing. Every downstream tool handles this format.

Paragraph-level

Timestamps appear once per paragraph or speaker turn.

[00:00:00] Speaker 1: Welcome back to the show. Today we're talking about pricing models for small businesses.

[00:00:18] Speaker 2: Thanks for having me. Pricing is one of those topics that...

Best for: podcast show notes, long-form reading, blog posts derived from interviews. The reader can find roughly where they are in the audio without a timestamp on every sentence cluttering the page.

Sentence-level

One timestamp per sentence.

[00:00:00] Welcome back to the show.
[00:00:03] Today we're talking about pricing models.
[00:00:09] Our guest has 20 years of experience.

Best for: search-driven review, legal or HR evidence transcripts, qualitative coding in research interviews. You can navigate to any sentence to verify exactly what was said. When unsure, sentence-level is the safe default: you can always strip the timestamps for display and keep them in the source file.

Word-level

A timestamp on every token.

[00:00:00.020] Welcome [00:00:00.380] back [00:00:00.640] to [00:00:00.780] the [00:00:00.880] show.

Best for: precise video editing, karaoke-style word highlighting, building a searchable audio index, ML training data. Word-level timestamps also let you clip a specific quote from a long recording without manually scrubbing the audio.

One precision note: not all word timestamps are equal. APIs like Deepgram return native word-level timestamps without a separate alignment pass. Native Whisper outputs segment-level timestamps (typically 5-30 second chunks) and interpolates word positions within each segment, which means errors accumulate over long files. Forced-alignment tools like WhisperX run a second phoneme-alignment pass after transcription and bring word accuracy to sub-100ms on clean audio. If word precision matters for your workflow, check how your tool produces those timestamps.

Subtitle-cue-level

Start and end timestamps in cue ranges, typically 1-7 seconds per cue, tuned for on-screen display.

1
00:00:00,000 --> 00:00:03,500
Welcome back to the show.

2
00:00:03,500 --> 00:00:09,200
Today we're talking about pricing models.

This is the SRT and VTT format. It is purpose-built for video rendering, not for reading. The SRT, VTT, and TXT export formats post covers the format specifics in detail, but the one rule to keep in mind: SRT cues cap at around 2 lines and 42 characters per line to stay within platform guidelines (Netflix, YouTube, broadcast). Words are grouped into cues by the transcription tool; you should not write cues by hand.

Use Case Lookup Table

Format Conventions

Three common formats exist, and mixing them is a recurring source of bugs:

• HH:MM:SS (e.g., 00:01:30). Simple, used in most paragraph and sentence transcripts.
• HH:MM:SS,mmm with a comma (e.g., 00:01:30,500). The SRT convention. Per the SRT format specification, the comma is mandatory.
• HH:MM:SS.mmm with a period (e.g., 00:01:30.500). The W3C WebVTT specification uses a period (U+002E FULL STOP) as the fractional separator. This is also the format most JSON APIs return.

If a file moves through multiple tools (export from a transcription service, import into a video editor, upload to a platform), normalize the decimal separator at the first step. Libraries that expect one form will silently fail or throw a parse error on the other.

Where Timestamps Come From

In AI transcription, timestamps originate from the speech recognition model. The key distinction is whether the model emits word timestamps natively or derives them from segment-level data:

• Native word timestamps (Deepgram and similar streaming APIs): the model returns a (word, start_time, end_time) tuple for each token directly.
• Segment-interpolated (native Whisper): the model produces segment boundaries and infers word positions, which can drift by hundreds of milliseconds on long recordings.
• Post-hoc forced alignment (WhisperX, aeneas, Montreal Forced Aligner): a separate phoneme-alignment pass maps words to audio after transcription, achieving sub-100ms accuracy.

Sentence- and paragraph-level timestamps are always aggregated from word-level data: the tool groups words by punctuation and speaker-turn pauses.

In human transcription, timestamps are inserted manually via a hotkey. Paragraph-level is common; sentence-level is slower and costs more; word-level is essentially never done by hand.

Common Editing Quirks

Drift on long files

For tools that use interpolated word timestamps, a 3-hour recording may show timestamps off by one or more seconds late in the file. The solution is a pipeline that re-anchors at silence boundaries, or use an API that emits native word-level timestamps from the start. If you are seeing drift consistently, check which timestamp method your tool uses.

Punctuation timing

Most tools attach a comma or period to the preceding word, so the timestamp is the end of that word. This is correct for reading but can cause an off-by-one issue if you programmatically slice audio at punctuation boundaries. Add a small buffer (50-100ms) when auto-clipping at punctuation.

Speaker turn boundaries

When Speaker 1 stops and Speaker 2 starts, the timestamp on Speaker 2's first word marks when their speech actually begins, not the start of the gap. For evidence clips or quote extraction, start the clip about half a second before the speaker label to include natural context.

Frame-rate alignment

For video work, timestamps need to align to the frame rate (24, 25, 30, or 60 fps). A timestamp like 00:01:30.123 does not map exactly to a frame at 30 fps. Most video editors round to the nearest frame, which is usually fine, but be aware that "click to seek" in an editor may land one frame off from the intended cut point.

Editing While Keeping Timestamps Intact

The hardest part of working with timestamped transcripts is editing text without breaking time anchors. Three approaches:

1. Use an editor that maintains alignment. Dedicated transcript tools (Otter, Descript, and others, per vendor documentation) automatically shift timestamps when you delete or insert words.
2. Edit text and re-align via forced alignment. Run aeneas or Montreal Forced Aligner on the edited text against the original audio. Accurate but adds a processing step.
3. Edit text and accept small drift. For paragraph-level timestamps, minor text edits rarely move the anchors enough to matter. Acceptable for show notes; not acceptable for legal transcripts.

For heavy edits (cutting half the content for a highlight reel), use a timeline-aware editor. Realigning word-level timestamps manually after large cuts is not practical.

Workflow: Podcast Show Notes With Chapter Markers

A common podcaster workflow:

1. Transcribe the episode with sentence-level timestamps.
2. Skim the transcript and pick 5-10 topic transitions (e.g., "How they started the company", "The pricing mistake").
3. Use the timestamps to write chapter markers: ID3v2 CHAP frames in MP3 files, MP4 chpl atoms in M4A files, or the Podcasting Index JSON chapter format in the RSS feed.
4. Publish show notes with timestamps as anchor links to the transcript page.

For a clean starting point, the /tools/audio-to-text tool on ConvertAudioToText produces a sentence-level transcript you can use for step 1 without a signup.

Workflow: Video Captions

For YouTube, TikTok, Instagram Reels, and broadcast platforms:

1. Transcribe the video with word-level timestamps.
2. Group words into subtitle cues within the 1-7 second window, respecting the 42-character-per-line limit.
3. Export as SRT (comma decimal) for broad compatibility, or VTT (period decimal) for web players.
4. Upload to the platform.

See SRT vs VTT vs TTML formats for a side-by-side on when each format is required versus optional.

FAQ

Do I always need timestamps in a transcript?

No. If you plan to read the transcript as a document, share it as meeting notes, or feed it to a summarizer, timestamps add visual clutter without benefit. They matter when you need to navigate back to the audio or sync text to video.

What is the difference between SRT and VTT timestamp formats?

SRT uses a comma as the decimal separator (00:01:30,500) while VTT uses a period (00:01:30.500). Both represent milliseconds. The distinction matters when tools process files programmatically: some libraries reject the comma form, others reject the period form. Normalize early if you pipe files between tools.

Why do word-level timestamps drift on long recordings?

Native Whisper emits segment-level timestamps (typically 5-30 second chunks) and interpolates word positions within each chunk. Small errors compound across chunks, producing drift of hundreds of milliseconds by the end of a long file. Forced-alignment tools like WhisperX re-anchor words to audio phonemes after transcription, bringing accuracy to sub-100ms. APIs like Deepgram emit native word-level timestamps without an interpolation step.

How do I edit a transcript without breaking the timestamps?

Use an editor that maintains alignment (most dedicated transcript tools do this). For light edits at paragraph-level granularity, small text changes rarely move the time anchors enough to matter. For heavy cuts, use a timeline-aware editor rather than a plain text editor, because manual realignment of word-level timestamps after large edits is very slow.

Sources

• W3C WebVTT specification: https://www.w3.org/TR/webvtt1/
• WhisperX paper (time-accurate long-form ASR): https://arxiv.org/html/2303.00747v2
• Deepgram Nova-3 documentation and word timestamps: https://developers.deepgram.com/docs/models-languages-overview
• Capital Captions subtitle guidelines (character limits, cue duration): https://capitalcaptions.com/services/subtitle-services-2/capital-captions-standard-subtitling-guidelines/
• Auphonic on MP3 ID3v2 chapter marks: https://auphonic.com/blog/2013/07/03/chapter-marks-and-enhanced-podcasts/
• aeneas forced aligner: https://github.com/readbeyond/aeneas