Digits or Words in an English TTS Script? Same Captions, One Time Error

No affiliate links in this post. This is our own production log from October 23, 2026 (KST); no tool vendor provided or reviewed it. The voice clips described here are AI-generated and were not published.

Short answer from one small test: in English, writing the numbers out as words in a TTS script did not change what our local auto-captions showed. Both versions came back with the same caption text, digits included. The one caption error was the same in both: a clock time, written 11:40 p.m. in our script, came back as 11.40 p.m.. So check times and other formatted numbers in the captions, whichever way you write them in the script.

The question we were checking

On October 22 we logged a Korean narration check: numbers spelled out in words came back as correct digits in the captions, but an ordinary noun was merged wrongly. One open limit there was that we had not tried English. ElevenLabs' help page on numbers and dates (re-read on October 23, 2026) recommends writing numbers, dates and symbols fully in words. We wanted to know what that choice does to the captions of a short English narration.

What we ran

One script, two versions. The two samples used different number spellings and different voice-conditioning setups, so they do not isolate the effect of spelling alone.

  • A (digits): "In 2023, a NOAA dive team filmed a golden orb at 3,250 meters. We rendered draft 2.5 of this clip at 11:40 p.m., cut 15% of the pauses, and kept the final cut under 30 seconds."
  • B (words): the same sentences with "twenty twenty-three", "three thousand two hundred fifty", "two point five", "eleven forty", "fifteen percent" and "thirty".

The second sentence is test wording made to hold a decimal, a clock time and a percent. It does not describe a real render.

Voice: VoxCPM2 run locally on CPU (cfg 2.0, 10 inference steps). Version A used a text voice description; that output was then used as the reference voice for B, so the two clips are close in voice but not the same in every way. Captions: whisper-large-v3-turbo run locally through mlx-whisper, language set to English, no editing. One render and one caption pass per version.

What came back

  • Clip length: A 12.64 seconds, B 15.04 seconds.
  • Captions for A and B were word-for-word identical: "In 2023, a NOAA dive team filmed a golden orb at 3,250 meters. We rendered draft 2.5 of this clip at 11.40 p.m., cut 15% of the pauses, and kept the final cut under 30 seconds."
  • The year, the depth with its comma, the decimal, the percent sign and "30" all matched the digit script.
  • The clock time was wrong in both: 11.40 instead of 11:40.

Why we read it this way, and what it does not show

Since the caption model wrote digits for both clips, writing words in the script did not keep words in the captions. If your caption style needs "three thousand" in words, you will have to edit captions either way. The time error is probably a caption formatting problem rather than a voice problem, because it was the same for two scripts written differently; we did not confirm that by listening.

Limits we did not test:

  • One script, one render and one caption pass per version. That is not a sample size you can generalize from.
  • The captions do not prove how the voice said the numbers. For example, a caption model may write "2023" for both "twenty twenty-three" and "two thousand twenty-three". Nobody listened to these clips specifically for this test.
  • The two clips had slightly different voice conditions, as described above.
  • Only one TTS model and one caption model, both local. We did not test ElevenLabs or its text normalization setting here.
  • The B clip is about 2.4 seconds longer, but its text is also longer. We did not compare pacing.

How to repeat it on your own script

  1. Pick the script lines that hold numbers: years, sizes, decimals, times, percents.
  2. Render them twice, once with digits and once in words, with the same voice settings.
  3. Run the same caption pass on both and compare each number against your script, character by character.
  4. Pay extra attention to clock times and other formatted numbers. In our run, those were the part that broke.
  5. Listen to the number lines yourself before publishing. The captions cannot tell you how a number was spoken.

Choosing a voice tool for your next script?

Our separate VoxCPM2 and ElevenLabs comparison covers a different production test and questions to consider before choosing a free local tool or a paid service. That separate guide contains a disclosed affiliate link. ElevenLabs was not tested in this digits-versus-words experiment; these caption results do not establish which voice tool is better.

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