waxcut

Frame-accurate, lossless MP3 splitting and duration parsing in pure Python with no ffmpeg, no subprocess, no decode step.

Cuts are made by parsing the file's own MPEG frame headers and byte-copying whole frames: output is bit-identical to the source, just shorter.

Install

pip install waxcut
# or
uv add waxcut

Usage

Quick duration check from the shell, no script needed — replace 'song.mp3' below (keep the quotes) with the path to your own file and run it as-is:

python -c "from pathlib import Path; from waxcut import load_audio_stream as l; print(round(l(Path('song.mp3')).playable_duration_ms / 1000, 1), 's')"

For actually splitting a file, here's the full pattern — load it once, then cut at whatever timestamp you want:

from pathlib import Path
from waxcut import load_audio_stream, frame_index_at, slice_bytes

stream = load_audio_stream(Path("song.mp3"))
print(f"{stream.playable_duration_ms / 1000:.1f}s")

# Split at the 90-second mark
cut_at = frame_index_at(stream.frames, target_ms=90_000)
first_half = slice_bytes(stream.data, stream.frames, 0, cut_at)
second_half = slice_bytes(stream.data, stream.frames, cut_at, len(stream.frames))

Path("part1.mp3").write_bytes(first_half)
Path("part2.mp3").write_bytes(second_half)

Splitting into more than two parts — split_at/join_frames collapse the loop above into one call:

from waxcut import load_audio_stream, split_at, join_frames, slice_bytes

stream = load_audio_stream(Path("mixtape.mp3"))
parts = split_at(stream, timestamps_ms=[90_000, 180_000, 270_000])

for i, part in enumerate(parts):
    Path(f"part{i}.mp3").write_bytes(part)

# join_frames is the inverse: reassembling parts reproduces the original
assert join_frames(parts) == slice_bytes(stream.data, stream.frames, 0, len(stream.frames))

Why not ffmpeg or a decode/re-encode library?

MP3 frames are self-describing, so their boundaries can be found directly from the byte stream — no decode step, no re-encode step, no external binary to shell out to.

waxcut also handles the parts that make naive frame-splitting subtly wrong:

• Skips leading ID3v2 tags when scanning for the first frame.
• Excludes the Xing/Info/VBRI VBR header frame — encoder metadata, not audio, and including it corrupts both output and duration.
• Parses LAME's gapless delay/padding extension, so reported duration matches what a real player shows, not just the raw frame count.

Duration parsing is cross-validated against mutagen's independent implementation to the millisecond (see Testing).

Scope

Parses MPEG-1/2/2.5 Audio Layer III — what "MP3" actually means. Layer I/II frames raise UnsupportedMp3Error rather than being silently mishandled, since virtually no real-world "MP3" file uses them.

Testing

uv sync
uv run pytest tests/ -v

Validated against mutagen's independent parser (duration must match exactly, including LAME gapless delay/padding) across CBR/VBR, mono/stereo, and multiple encoder tags. Where ffmpeg/ffprobe are available, every split output is independently decoded to confirm it's valid. Fuzzed continuously with ClusterFuzzLite.

Contributing

Bug reports and pull requests are welcome — see CONTRIBUTING.md for the dev setup and PR process. Report security vulnerabilities per SECURITY.md rather than as public issues.

License

Apache-2.0 — see LICENSE.