[devlato]

Songs can be exported in every popular format, including the ones the web plays, and every song folder gets its song.mp3 without ffmpeg.

claudius export <song folder or audio file> -f mp3,flac:16,opus writes the master as:

• WAV: 24-bit, 16-bit at 44.1 kHz (CD), or 32-bit float;
• AIFF, FLAC, Apple Lossless;
• MP3 at 320 to 128 kbps, or VBR V0 to V9;
• AAC in .m4a, Opus in .opus, and Ogg Vorbis.

--list shows every format, and any this machine cannot write. The studio menu, MCP tool and HTTP download come in a follow-up pull request.

How it's built

A new crate, claudius-encode, the counterpart of claudius-decode:

• In Rust, in process:
  • WAV (hound) and AIFF (written here);
  • FLAC (flacenc, the pack installer's code moved here);
  • MP3 (rusty_mp3) and Opus (opus-pure).
• Built in: Ogg Vorbis, through vorbis_rs, which compiles libvorbis with aoTuV via cc.
• The system's encoder: AAC and Apple Lossless come from Core Audio's afconvert on macOS, and from ffmpeg elsewhere when it is installed.

Licences: Apache-2.0 and BSD-3 only. LAME (LGPL) was ruled out by deny.toml. A pure-Rust AAC encoder exists (oxideav-aac), but it took 36 s to encode 30 s of audio.

What else is in it:

• Resampling to 44.1 kHz: a polyphase Kaiser-windowed sinc with about 95 dB of alias rejection, flat to 20 kHz.
• 16-bit files: TPDF dither from a fixed seed, so the same song always makes the same file.
• Tags: ID3v2.4 in MP3, Vorbis comments in FLAC, Opus and Ogg, and ffmpeg's tags for AAC and ALAC.
• FLAC: every file is decoded again before it is kept. Two ways of choosing predictors are tried and the smaller valid one wins. A 24-bit master went from 97% of its raw size to 82%.
• Supporting changes:
  • claudius-decode gains decode_bytes;
  • the engine gains an export module;
  • song.mp3 and the server's playback MP3 are encoded in process;
  • the doctor's ffmpeg line now says what ffmpeg is still for;
  • the telemetry Worker learns the export command. It needs a redeploy after the merge.

How I checked

Quality, on 30 s of a real render, against ffmpeg's encoders at the same bitrate (mean difference in third-octave band energy, decoded by ffmpeg):

	claudius	reference
MP3 320	0.26 dB	LAME 0.39 dB
MP3 V0	0.47 dB	LAME 0.48 dB
MP3 192	1.28 dB	LAME 1.28 dB
Opus 160	0.67 dB	libopus 0.71 dB
AAC 256	Core Audio, the same encoder	Apple 0.19 dB

Every format of a 3.5-minute song, all ten exported in 4.4 s, decodes:

• Core Audio decodes all ten, with exact lengths, so AAC and Opus playback is gapless.
• symphonia decodes all but Opus, which it does not support.
• Chromium 156 decodes MP3, AAC, Opus, Ogg Vorbis, FLAC and WAV.
• WebKit 27.2 decodes all ten. The ALAC file, which Apple's own encoder wrote, stops at 158 s in WebKit. AIFF and ALAC are documented as formats for Apple's apps, not the web.

Tests:

• new unit tests: specs, lossless WAV and AIFF round trips, dither, FLAC tags and the bound, the MP3 Info frame, gapless Opus, deterministic Vorbis, Core Audio AAC and ALAC, and the resampler's alias rejection;
• the CLI test exports a rendered song to three formats and decodes them, and checks the errors;
• cargo xtask ci, cargo deny check, cargo xtask site with lychee (0 broken links), and the Worker's tests pass.