ademvessell/arkhe-lyric-lock — explained in plain English
Analysis updated 2026-05-18
Generate word-level lyric timing for a song to import as SRT subtitles into a video editor.
Catch sung sections that are missing from your lyric sheet before shipping a lyric video with gaps.
Manually fix a handful of low-confidence word timings in the browser-based repair bench.
Compare CTC forced alignment against a transcription-based approach for timing accuracy on sung audio.
| ademvessell/arkhe-lyric-lock | 0xallam/my-recipe | 0xhassaan/nn-from-scratch | |
|---|---|---|---|
| Stars | 0 | — | 0 |
| Language | Python | Python | Python |
| Last pushed | — | 2022-11-22 | — |
| Maintenance | — | Dormant | — |
| Setup difficulty | moderate | moderate | moderate |
| Complexity | 3/5 | 2/5 | 4/5 |
| Audience | developer | general | developer |
Figures from each repo's GitHub metadata at analysis time.
First run downloads about 1.2 GB of aligner model weights, and demucs is recommended in its own virtual environment.
Arkhe Lyric Lock is a tool for timing lyrics to a song so they can be used in a lyric video, and its main selling point is that it tells you when it is not confident in its own results instead of quietly guessing. You give it an audio file and a text file of the lyrics, and it works out exactly when each word is sung. The README explains that this is different from typical speech transcription tools, which try to guess what words were sung, a much harder task on singing than on speech, since held notes and dense music mixes trip them up. Instead, this tool already knows the words from your lyric sheet and only has to figure out when each one happens, which the README claims produces much more accurate timing, backed by a small comparison table against a transcription-based approach. Under the hood, it first separates the vocal track from the rest of the song using a tool called Demucs, then aligns the lyrics to that isolated vocal using a technique called forced alignment, then makes a second pass to fix timing on held notes. Every placed word gets a confidence score, and the tool will flag or even stop entirely if it hears singing that is not in your lyric sheet, rather than showing a blank or wrong result. Repeated lines already in the sheet are handled automatically, and unclear ad-libs are left out rather than forced to match text that does not fit. The output includes a timed data file, subtitle files that import into video editors like Premiere or Final Cut, and a plain reference video. A companion repair tool is a webpage you open directly in your browser, with no server needed and nothing sent off your machine, where you can review and manually fix any words the aligner flagged as uncertain. The README is upfront about limits: it needs the lyric text ahead of time since it cannot transcribe singing on its own, very fast rap can still be shaky, and background vocals are intentionally skipped. The project's own code is released under the MIT license, but it depends on a separate speech alignment model whose weights carry a non-commercial license, so the README points to a notice file before anyone considers commercial use.
A local tool that times lyrics to a song for lyric videos, flagging any words it isn't confident about instead of guessing.
Mainly Python. The stack also includes Python, PyTorch, Demucs.
The project's own code is MIT and free to use, but it relies on a separate alignment model whose weights are non-commercial only, so commercial use needs a substitute model.
Setup difficulty is rated moderate, with roughly 1h+ to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.