Auto-Captions Got the Numbers Right and One Noun Wrong: A Korean Narration Check

No affiliate links in this post. This is our own production log from October 22, 2026; no tool vendor provided or reviewed it.

If you write a narration script with the numbers spelled out as words, local auto-captions can still turn those numbers back into clean digits. In our test, the numbers were fine. The word that broke was an ordinary noun with a grammar particle attached, and the caption would have shipped wrong if we had only checked digits.

The problem we were checking

We make short Korean science videos with an AI narration voice, and the subtitles are burned in from a speech-to-text pass over the final audio. After an earlier test where a transcript got a date wrong, we now write years and depths out in words in the script, because text-to-speech tools are often unsure how to read raw digits. The question this time was simple: when the script says the numbers in words, what do the auto-captions show, and where do they go wrong?

What we actually ran

  • Script: one 4-sentence Korean narration, 153 characters, written by us. It contains a year and a depth spelled out in Korean words (two thousand twenty-three; three thousand two hundred fifty meters).
  • Voice: a local VoxCPM2 run with a reference voice we generated ourselves. One completed synthesis was used (an earlier run was stopped to fix a sentence in the script before it finished).
  • Captions: a local speech-to-text pass over the finished 21.47-second audio, split into 12 caption chunks, with 0 dropped segments.

What the captions said

  • Year: spoken as words, captioned as 2023년. Correct.
  • Depth: spoken as words, captioned as 3250m. Correct.
  • One noun: the script says 말미잘이 (sea anemone, plus the subject particle). The caption came back as 말미자리, which merges the noun and the particle into a word that is not in the script. In the burned-in subtitle that would read as a different, meaningless word right at the reveal line of the video.

Our automatic sentence check scored the final narration at 0.866 coverage against the script, comfortably above our 0.5 minimum. That score did not flag the wrong noun. A coverage number tells you sentences are present, not that every word is right.

What we decided and why

We corrected the caption text to the authored script word 말미잘이 before any final review, and kept the original audio unchanged. We did not regenerate the voice: the script was right, and we had no evidence the voice said the word wrong. Afterwards, the full final audio was listened to once by an AI review tool in our separate review step; this post does not claim that one AI listening pass proves the speech was perfect, and no person listened specifically for this test.

The decision rule we are keeping: compare captions against the authored script, and give extra attention to names, technical nouns and particles next to them, not only to dates and numbers.

Limits of this result

  • One script, one completed voice run, one caption pass. This is a single case, not a rate.
  • We cannot say whether the voice blurred the word boundary or the speech-to-text model mis-segmented it. We only know the caption text was wrong.
  • Korean particles attach directly to nouns, so this kind of merge may be more common in Korean than in English. We did not test English here.
  • We did not compare caption models, settings or speeds, and we make no quality ranking.

How to repeat this check on your own narration

  1. Write numbers, dates and units out in words in the script you send to the voice. ElevenLabs' own help page, which we re-read on October 22, 2026, gives the same advice: it recommends "writing numbers, acronyms, dates and symbols fully, in words".
  2. Generate the final audio once and keep the authored script as the reference text.
  3. Run speech-to-text on the final audio, not on an earlier draft.
  4. Diff the caption text against the script word by word. Look first at digits, then at proper names and technical nouns, and at whatever grammar sits right next to them.
  5. Fix captions to the authored script when the script is right; only re-record the voice if a listen confirms the speech itself is wrong.

For the earlier case where both transcripts contained every sentence but one date was off, see our October 13 test: Two TTS transcripts disagreed on a date.

Sources checked

  • ElevenLabs Help Center, "Why are numbers, dates, symbols and acronyms not properly pronounced or spoken in the correct language?" (re-read October 22, 2026): elevenlabs.io help page
  • Our own production records for this video (script, caption file, caption-correction note), October 22, 2026. The video itself is not published yet.

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