LRC lyrics file from an AI music track: Sume STT segments in Python

Why an LRC file

YouTube's subtitle help page lists .lrc among the supported subtitle and caption file types. LRC is a line-per-lyric format used for synced lyrics, with a bracketed time before each line.

Sume's Music Router result carries the lyrics in result.lyrics, either as a section map or as lyrics text, but those lyrics have no times. To get times, run the finished track through STT 1.0 with sentence segmentation.

Two sources of lyrics, and the caveat

You have two candidates for the text: the lyrics the music job returned, or what STT hears. They differ. STT transcribes what it can recognise, and singing over instruments is harder than a spoken voice, so expect misses, repeated lines and mistakes.

That makes this a draft generator, not a finished product. Treat the output as a starting file you proofread against the real lyrics. I have not verified STT accuracy on sung vocals, so check every line.

The script

The script converts segments into LRC lines using start times. Replace the sample with your job's segments.

def stamp(sec):
    cs = round(sec * 100)
    m, cs = divmod(cs, 6000)
    s, cs = divmod(cs, 100)
    return f"[{m:02d}:{s:02d}.{cs:02d}]"

segments = [
    {"text": "City lights are calling", "start": 12.4, "end": 15.9},
    {"text": "Hold on, don't let go", "start": 15.9, "end": 19.3},
]
lrc = "\n".join(stamp(s["start"]) + s["text"] for s in segments)
open("song.lrc", "w", encoding="utf-8").write(lrc + "\n")
print(lrc)

Making it better

• Replace each recognised line with the matching line of the real lyrics, keeping the time.
• Delete lines STT produced for instrumental sections.
• Listen with the file in a player that supports LRC before uploading.
• Keep the track as a wav when you can, since compressed audio is a worse input for recognition. See wav vs mp3.

Using the generated lyrics as the text

A better workflow keeps your intended lyrics as the text and uses STT only for times. Match each lyric line to the nearest recognised segment by order, and copy the segment start time to the line. Where counts differ, align by hand.

This is slower than accepting recognition output but gives a file that reads correctly. For a one-off release it is worth the ten minutes.

Cost

The music generation is a fixed $0.125 per job and STT is $0.01 per audio minute, so a 3-minute track costs about $0.03 to transcribe. The API pricing page holds the live book.