Hallelujah vs The Competitor: Why ChangeLyric Won
A customer hired both ChangeLyric and a competitor to rewrite Leonard Cohen's classic. Here's why ours was 'much better.'
Posted by
Related reading
Get Up 10 (Cardi B) - Custom Personal Lyrics
Listen to a custom version of Cardi B's Get Up 10 rewritten as a personalized hardship to triumph story. Before and after audio from ChangeLyric.
Proud of Your Boy (Aladdin) - Mother Son Wedding Dance
Listen to a custom version of Aladdin's Proud of Your Boy flipped from apology to celebration. Before and after audio from ChangeLyric.
Special Death (Mirah) - English Lyric Adaptation with Vocal Pacing
Mirah's Special Death adapted with custom English lyrics that fit the original melody, plus a frank look at why ChatGPT-written revision notes do not help on audio deliveries.

Most customers do not tell you when they hire the competition. Arthur did. He came to ChangeLyric with a request to rewrite Leonard Cohen's "Hallelujah" (in the Jeff Buckley style) as a personalized tribute for someone named Viki. When our first version did not match Buckley's voice closely enough, he hired AI Music service for comparison. His verdict: our version was "much better."
The Request
Arthur wanted a complete rewrite. "Hallelujah" became "Vikidoo." The lyrics tell a life story: meeting Helene in Paris, building a career at KKR in London, the hobbies (skiing, tennis, golf), the personality quirks (Tabasco in every suitcase), and the signature move (playing guitar badly after a few drinks). Arthur provided detailed notes including pronunciation guides (Helene with a French accent, "Vikidoo" as "Vicky doo") and specific instructions about where the new lyrics should hit in Buckley's vocal arc.
The challenge: Jeff Buckley's voice is notoriously difficult to replicate. His "Hallelujah" is not just a vocal performance. It is a masterclass in restraint and release, whispered verses exploding into the chorus. Most AI vocal systems struggle with this dynamic range.
Listen & Compare
Hear the original song and the custom version side by side
Transcripts are auto-generated and may not perfectly reflect the audio.
“Hi Will. Thanks for this! Honest feedback: I was hoping the voice would be much closer to the original. Because it wasn't I also tried AI Music service. And your version was much better so I would have used yours. But in the end I decided to ask my colleague who is a singer and she did a great job. The AI service really only makes sense to me when you're able to fully copy the original. In this case Jeff Buckley i guess is a very difficult voice to replicate. Am sure we are not far away.”
— Arthur H.
The Competitor Test
Arthur's review is unusual because it names the competition. He tried AI Music service when our first pass did not nail Buckley's voice. His comparison was direct and unambiguous: "your version was much better." This matters because it confirms something we observe internally: for difficult vocal sources (and Buckley is among the most difficult), our engine produces more usable results than alternatives.
Hear it for yourself. This is the version Arthur received from the competing AI Music service after our first pass did not satisfy him.
The Competitor's Version (AI Music Service)
Compare it against our delivered version in the player above. Both systems struggle with Buckley's unique vocal texture. Neither fully captures his voice, but ours was close enough that Arthur would have used it.
Why Both Versions Fall Short of Buckley
There is audible autotune in the competitor's version, most obvious on the sustained notes. Both their version and ours almost certainly use an RVC model trained on Buckley's voice. RVC is good at imprinting a target singer's timbre onto an existing vocal, but it cannot invent performance nuance the source vocal does not already have.
Buckley's Hallelujah is a raw, almost unplugged take. The performance lives in the micro details: the breath catches, the slight pitch drifts, the held vibrato that decays into silence. An RVC pass over a clean session vocal flattens those into something polished, which is why the competitor's version sounds technically correct but emotionally flat.
The two pipelines also differ in how much is AI generated. AI Music Service's result was likely sung by a session vocalist and then converted through RVC, which takes less total work. Ours used multiple AI generated vocal stems engineered to approximate Buckley's phrasing, then a final RVC pass to lock the timbre in.
That approach is more labor intensive but it lands closer to the source. We use our own RVC tooling for projects like this. The ChangeLyric Voice Changer ships over 200 curated models plus custom voice training, and it handled the final pass on this song.
Why It Worked
The final delivery preserves the structure of Buckley's arrangement. The verses maintain his conversational phrasing. The chorus builds to the "Vikidoo ya" release at exactly the moments Buckley would have hit the high notes. Even the guitar is there, that famously wrong cheerful strumming Arthur wrote into the lyrics.
Arthur ultimately went with a live singer. For this specific use case, that made sense. Human performers have advantages AI cannot yet match. But when he needed an AI option, he preferred ours. For Jeff Buckley, one of the hardest voices in popular music to replicate, that is the best outcome we can ask for.
Browse more projects like this in our lyric swap showcase, or start your own custom lyric swap.