Both TTS Files Had All 4 Sentences, but One Date Was Off: A Korean Script Test

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. The test below used our own accounts and our own files.

A text-to-speech file can contain every sentence of your script and still get a date wrong. On October 13, 2026 we put one short Korean script, written by us, through a local VoxCPM2 run and a cloud ElevenLabs voice, then transcribed both files end to end. Both transcripts contained all four sentences. They did not agree on the date, and that is the only reason this post exists.

The script and what we ran

The script is 130 Korean characters and four sentences. In English it says: "Today we make the narration for the video. We compare fourteen items priced at 2,480 won, from January 3 to December 31. Prices are read as they are, not rounded to the nearest thousand. Finally, the file is saved as final_v2.mp4."

It is built on purpose: a date range, a price, a count and a file name. Those are the tokens where a wrong word changes the meaning of a video.

  • Local: VoxCPM2, unconditioned voice, one generation. The file is 17.28 seconds.
  • Cloud: ElevenLabs web app, a stock voice already in our account, no cloning, one generation click, variation 1 of the 2 shown. The file is 23.07 seconds.
  • Check: full-file Whisper large-v3-turbo transcription (language set to Korean), run locally on both files, raw output kept.

What the transcripts said

  • Local file: all four sentences present. Date range read back as January 3 to December 31, price 2,480 won, count 14. The file name came back as "Final V-Toin MP4", which is not final_v2.mp4.
  • Cloud file: all four sentences present. Price and count matched. The end date came back as December 30, not December 31. The file name came back as a spoken "final underscore v2.mp4".

What we do not know

We did not listen to either file end to end, so we cannot tell a spoken error from a transcription error. The December 30 might be the voice, or it might be Whisper. The "V-Toin" might be a mispronounced "v2" or a recognition miss on an underscore. This is one script, one run per engine, two different voices, and different codecs and lengths. The 5.8-second length difference says nothing about speed or quality, and we are not ranking the two.

What we changed because of it

We stopped treating "all sentences are present" as a pass. For any narration we now mark dates, prices, counts and file names, and check those tokens against the transcript first, then listen to exactly those seconds. A transcript match is still not a listening check, so the listening step stays.

ElevenLabs' own help article on this problem says numbers, dates, symbols and acronyms can be mispronounced and recommends writing them out in words; it also describes an API text-normalization setting with on, off and auto. That is the next thing we will test on this same script, with the same transcription check. We have not run it yet, so we do not know whether it fixes the date.

Try it on your own script

  1. Write a short script with one date range, one price, one count and one file name.
  2. Generate it with each engine you are considering and keep the raw files.
  3. Transcribe the full files and compare only those four tokens against the script.
  4. Listen to the seconds around any mismatch before you blame the voice.

If you want to try the cloud side, this is our referral link: DigitalHarvest (NSST) is an independent affiliate of ElevenLabs and may earn a commission if you subscribe through this link: ElevenLabs. The original audio files and reproduction commands are in our Korean write-up of the same test: smartaireviewer.com (Korean).

Earlier posts from the same log: VoxCPM2 vs ElevenLabs on the same 30-second script and One day of video-credit logs.

Sources checked

Affiliate relationship and trademark attribution

ElevenLabs and the ElevenLabs logo are registered trademarks or trademarks of ElevenLabs, Inc. Used with permission. DigitalHarvest (NSST) is not sponsored by, endorsed by, or affiliated with ElevenLabs except as an independent participant in the ElevenLabs Creator Affiliate Program.

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