I have some songs in English that do not have published lyrics due to those are not popular. So I'm looking for an online service or a standalone app that can extract (recognize) lyrics (as text) from a song (say, mp3).
3
-
1I don't think that speech synthesis is at that advanced level yet, especially when there are other sounds (like music) involved. Try uploading the file to YouTube and take a look at the closed captions to get an idea– TymricCommented Oct 2, 2014 at 21:25
-
Keep in mind that it is impossible to get 100% accuracy in this as in singing things like rhythm and amount of syllables are often more important than "correct" words. So, words will be cut of or at a grammatically incorrect place or intentionally misspoken to add multiple meanings etc. understanding art is something computers perform horribly at.– Angelo FuchsCommented Oct 8, 2014 at 10:21
-
Melobytes is decent but not perfect– IknowItCommented Jun 26 at 14:09
Add a comment
|
1 Answer
As speech recognition has advanced much in the past few years, I can now provide a local and FOSS solution at this time: OpenAI Whisper. You can also use this Google Colab Notebook if it is not convenient for you to install it locally. An improved version to Whisper is WhisperX, which features faster transcription, accurate word-level timestamp, and transcription with multiple speakers.
There is also a qualitative experiment of using Whisper to transcribe songs of various difficulty levels for transcription: Zero-Shot Song Lyrics Transcription Using Whisper. Also: Paywall-free version. You can see that it can transcribe songs up to the difficulty of Amon Amarth's The Pursuit of Vikings with only some single-word errors. Although from my personal experience, the vocal has to be at least kind-of-intelligible for humans, in order for for Whisper to transcribe it. I have tried transcribing a really raw and unintelligible black metal song with Whisper and it doesn't work in that case.
-
I was able to use this on my song collection, and it seemed to be a pretty good job. However, there were still some failures, and I have NOT had time to do a word for word comparison to the audio. Commented Oct 3 at 18:23