Before You Post an AI Song: Check Lyrics, Clipping, and the Ending
A correct caption does not prove an AI singer sang the correct words. A file that opens without errors can still have unclear lyrics, overloaded peaks or an awkward ending. Check those separately before choosing an export to upload.
We reviewed three Gemini Lyria 3.5 song files made with new, AI-assisted Korean lyrics on September 14, 2026. Our evidence was local speech recognition, audio measurements and selected final-video frames. We did not complete a human listening review, and none of the three songs passed our publication check. These are workflow examples, not a benchmark of the music generator.
What the three files actually showed
| Check | v1 | v2 | v3 |
|---|---|---|---|
| Actual length | 32.60s | 61.96s | 63.01s |
| Exact lyric lines | 2/6 | 2/6 | 2/6 |
| Lines with any corresponding text | 6/6 | 6/6 | 3/6 |
| True peak, dBTP | 0.0 | +0.1 | −0.1 |
“Exact” ignores spaces and punctuation. A focused v2 check recovered another line, but two key nouns stayed uncertain. Version three used rewritten lyrics. Different wording and recognition windows make these counts unsuitable as a controlled quality ranking. Missing transcript text also does not prove a singer omitted a line.
Use transcription to find listening questions
Keep the expected lyrics separate. Run an initial transcript without feeding those lyrics into the recognizer, save the raw result, then compare line by line. We used a locally run MLX Whisper large-v3-turbo model. Record the model, settings and file identity so another check refers to the same recording.
In v2, the intended 퇴근했어 (“finished work”) became 개근했어 in both the main transcript and a short-window check. 편의점 (“convenience store”) became 변을 전 or 변의 전. These are recognition discrepancies, not confirmed pronunciation errors. Listen to the disputed phrases without watching captions, then compare with the lyric sheet. If the words remain ambiguous, revise the phrase or recording.
Check suspicious additions too. Our earlier transcripts invented repeated syllables over a digitally silent tail. Keep that raw evidence, flag the hallucination and exclude it from approved captions. Repeating recognition until it produces the expected answer is not verification.
Measure the file you will actually upload
Sample peak measures stored sample amplitudes; true peak estimates peaks between samples. Integrated loudness describes the program’s loudness level. The FFmpeg ebur128 documentation explains its scanner and oversampled true-peak mode.
With FFmpeg installed, run this read-only check, replacing the filename:
ffmpeg -hide_banner -nostats -i "final.mp4" -map 0:a:0 -af ebur128=peak=true -f null -Read the final summary’s integrated loudness and true peak. Run it separately on the source audio and final video. Record which file each result belongs to; an encoder setting is not a measurement of the encoded output.
Our v2 source measured −10.8 LUFS and about +0.1 dBTP. After trimming, fading and normalization, the final AAC measured −13.78 LUFS and −1.86 dBTP. Those values document more peak headroom in this export. They do not establish clear singing or the absence of every audible artifact. Our −14 LUFS/−1.5 dBTP render targets were editing choices, not YouTube requirements. FFmpeg’s loudnorm reference describes the separate loudness and peak targets.
61.96 seconds
+0.1 dBTP
Trim → fade → normalize
Export the video
27.60 seconds of audio
−1.86 dBTP
Separate decision: sung lyrics still unverified.
Finish the phrase, then inspect the ending
The v2 recognizer placed the final lyric near source time 27.84 seconds. We made a review draft from 2.7–30.3 seconds, leaving about 2.46 seconds after that estimated endpoint. The 27.6-second audio used a 0.85-second end fade. The afade reference documents fade start and duration.
The final audio’s last ten milliseconds measured about −60.7 dBFS RMS, but a quiet tail does not prove the melody resolves naturally. Listen through the last word and final beat. Also inspect the opening for a clipped syllable, and watch the last frame. Our draft remained unapproved because lyric clarity was unresolved.
Keep one release checklist
- Confirm every sung line by listening; log uncertain words.
- Measure the final export and inspect its opening and ending.
- Record the exact filename, hash, date and remaining issues.
- Check rights and the platform’s AI disclosure before uploading.
YouTube includes AI-generated music in its disclosure examples. Select the applicable AI-use disclosure during upload. Google’s consumer terms say it does not claim ownership of generated original content; that does not guarantee exclusive copyright. Content ID requires exclusive rights. Passing technical checks does not guarantee monetization or ownership.
For the final video files, our free exact-file checker can help distinguish an identical renamed copy from a changed export. It does not judge lyrics, sound quality or rights. NSST built that tool; this article has no affiliate links.
Comments
Post a Comment