Which Punctuation Makes a Longer TTS Pause? We Measured 8 Clips

No affiliate links in this post. This is our own production log from October 25, 2026 (KST); no tool vendor provided or reviewed it, and we did not test ElevenLabs or any hosted voice service here. The voice clips are AI-generated with a local model and were not published.

Short answer from one small test: in our local VoxCPM2 voice, an em dash (—) between phrases gave the longest pause in both renders, roughly 0.4 seconds at the first break. A comma and a period gave shorter breaks, and the period was not reliably longer than the comma. An ellipsis (...) was one of the least predictable: long in one render, short in the other. If you need an exact pause length, punctuation is not a timing control in this model.

The question

When an AI narrator rushes from one step to the next, the usual advice is "add a comma" or "use an ellipsis". Vendor help pages give different advice on which mark to use, and the local model we use for narration, VoxCPM2, gives no pause guidance in its model card. So we measured it on our own machine instead of guessing.

What we ran

  • Model: VoxCPM2 (openbmb, Apache-2.0), local CPU run, the same settings we use for our narration (cfg 2.0, 10 inference steps).
  • Voice: every clip was cloned from the same reference clip, our own AI narration from October 23. Only the text changed.
  • Script: one sentence, "Export the clip_ check the captions_ then upload it tonight." The two blanks were filled with a comma, a period, an ellipsis, or a spaced em dash. In the period version, "Check" and "Then" are capitalized, because that is how you would write it.
  • Repeats: each version was rendered twice, so 8 clips in total, rendered 21:12–21:16 KST.
  • Measurement: FFmpeg silencedetect (−40 dB, at least 0.05 s) found the silent stretches; local Whisper large-v3-turbo word timestamps told us which silence sits at each break. We report the longest silence at each of the two breaks.

Results

MarkRenderBreak 1 (clip → check)Break 2 (captions → then)Both breaksClip length
Comma10.09 s0.23 s0.31 s3.20 s
Comma20.22 s0.17 s0.39 s4.16 s
Period10.21 s0.17 s0.38 s3.36 s
Period20.11 s0.14 s0.25 s3.36 s
Ellipsis10.30 s0.24 s0.54 s3.52 s
Ellipsis20.18 s0.14 s0.32 s3.84 s
Em dash10.40 s0.22 s0.62 s3.52 s
Em dash20.43 s0.37 s0.80 s4.16 s

Clip length is the whole file, including silence at the start and end. Pause values are rounded from the raw measurements.

Three things stood out:

  • The em dash won both times. Its combined pause (0.62 s and 0.80 s) was longer than any comma, period, or ellipsis render.
  • Same text, different timing. The two comma renders came out 3.20 s and 4.16 s long. Render-to-render variation was about as large as the differences between some marks, so one render of each is not enough to choose.
  • Captions ignore your punctuation. Whisper wrote all 8 clips the same way, with commas: "Export the clip, check the captions, then upload it tonight." Writing a dash or an ellipsis into the script did not carry through to Whisper's transcript, so don't rely on auto-captions to keep it.

Why we chose the dash for our next narration

We want a clear beat between steps without the hesitant tone an ellipsis can add. In this test the em dash gave the longest break in both renders, and the most consistent first break (0.40 s and 0.43 s), and it costs nothing to try. ElevenLabs' own help page recommends something similar for its voices: it says a dash or em dash often works, and that an ellipsis "usually also adds some hesitation or nervousness to the voice". We did not test that claim ourselves.

What this does not show

  • Two renders per mark, one sentence, one cloned voice, one model. Treat the ranking as a hint for your own test, not a rule.
  • Silence below −40 dB is a signal measurement, not a listening test. Nobody rated how natural each pause sounded, and short stop-consonant gaps can also register as silence.
  • Whisper word timestamps are approximate, so the break windows are approximate too.
  • We did not test SSML, other local models, longer scripts, or any hosted voice service.

Repeat it on your own voice

  1. Write one sentence with two breaks and make four copies: comma, period, ellipsis, em dash.
  2. Render each copy at least twice with the same voice and settings.
  3. Run ffmpeg -i clip.wav -af silencedetect=noise=-40dB:d=0.05 -f null - and note the silence nearest each break.
  4. Pick the mark whose pause is long enough in every render, then listen to it before you use it.

If you need an exact pause length rather than a nudge, a hosted service with a timing tag is the other route. ElevenLabs documents an SSML break tag, <break time="1.0s" />, up to 3 seconds on Multilingual v2, Flash v2 and Flash v2.5. Its docs say Eleven v3 and v4 do not support the tag, and that too many breaks in one generation can cause instability. A local model plus a dash was enough for our short narrations.

Previous step in this series: Whisper wrote 11.40 p.m.: a style prompt didn't fix it, a regex did.

Sources checked

Comments

Post a Comment

Popular posts from this blog

FFmpeg Shows −1.5 dBTP. Did You Actually Fix the Audio?

An output_tp of −1.50 in an FFmpeg log does not mean your original audio has been fixed. A command can show that number without saving any corrected audio. Before uploading, measure the file you actually exported. This follow-up to our AI song quality checklist tackles one practical mistake: confusing filter output statistics with a saved deliverable. You need FFmpeg installed and a local audio file. These commands process files on your computer; they do not upload your recording. One recording, two different stages On September 18, 2026, we measured one newly generated, roughly 2-minute-20-second MP3. Its integrated loudness was −11.41 LUFS and its true peak was +0.84 dBTP . The same analysis log showed output_tp: -1.50 . input_i and input_tp describe the file supplied to that command. output_i and output_tp describe the signal after the filter in that execution. They do not establish the measured properties of a separately encoded final file. We then reduced gain by...
Read more

VoxCPM2 vs ElevenLabs for a 30-Second Science Short: We Timed Both on the Same Script

Image
Disclosure: DigitalHarvest (NSST) is an independent affiliate of ElevenLabs. We may earn a commission if you subscribe through the referral link in this article. ElevenLabs has not sponsored, reviewed, or approved this article. This relationship does not change the numbers below, and the free local option wins on cost in our own test. We make 30-second science Shorts from public-domain NOAA ocean footage, and every one needs about six lines of narration. On October 4, 2026 we took one finished script and voiced the same six lines twice: once with VoxCPM2, an open-source model running on our own Apple Silicon Mac, and once with ElevenLabs on the paid Creator plan we already use. Then we ran both outputs through the same Whisper check our pipeline uses before anything is published. Here is what came out. The script we used The episode is "Why doesn't hot water on the seafloor boil?" (a small black smoker vent filmed at 2,947 m by NOAA Ocean Exploration). Six lines, 428...
Read more

Before You Upload an AI Song or Video: The Checks We Run (Free Commands)

The FFmpeg and file-hash commands below run locally. No purchase is needed to follow them. In September 2026 we published a 140-second "music video" for one of our AI songs. It was one still picture with a small moving waveform along the bottom. Our pre-upload checks had measured the length, the resolution and the loudness. Nobody had checked whether the picture actually moved. We deleted it from our channel the next morning. This post shows the checks we run now. Every command below runs on your own computer with ffmpeg , and every number comes from files we measured ourselves. 1. Loudness and true peak (free, one command) ffmpeg -hide_banner -nostats -i song.mp4 -af ebur128=peak=true -f null - Read the Summary block at the end: I: is the integrated loudness in LUFS. Peak: under "True peak" is the true peak. ffmpeg prints the unit as dBFS. A test file we deliberately pushed too loud read I: -8.4 LUFS and Peak: 3.5 . Our own working baseline is ...
Read more