glint_audio_pure 0.6.1

Pure-Dart MPEG-1 Layer III (MP3) encoder — no native code, no FFI, runs identically on every platform including the web. A pure-Dart sibling of the FFI-backed glint_audio, ported from the glint codec suite.

glint_audio_pure

A pure-Dart MP3 (MPEG-1 Layer III) encoder — no native code, no FFI, no runtime dependencies. It runs identically on every platform Dart targets, including the web, where FFI-backed codecs can't go.

It is the pure-Dart sibling of glint_audio (FFI bindings to the glint C++ codec suite). Use glint_audio when you want the full suite (MP3/AAC/Opus/WAV + resampling) at native speed; use glint_audio_pure when you need MP3 encoding that just works everywhere with zero setup.

import 'dart:typed_data';
import 'package:glint_audio_pure/glint_audio_pure.dart';

// pcm: mono samples in [-1, 1]
final Uint8List mp3 = mp3EncodeMono(pcm, sampleRate: 44100, bitrate: 128);
// -> a standard .mp3 any decoder plays.

Why

• All platforms, including web. Pure dart:core / dart:math / dart:typed_data. No dart:ffi, no bundled binaries, no build hooks.
• Standard output. Produces conformant MPEG-1 Layer III that ffmpeg, browser <audio>, and every other decoder play.
• Faithful port. Clean-room port of glint's encoder. The DSP front-end is machine-equivalent to the C++ reference (subband 5e-15, MDCT 7e-16 relative error) and the quantizer reproduces glint's per-granule decisions.

What it does

The full MPEG-1 Layer III encode pipeline:

1. 32-band polyphase (subband) analysis
2. MDCT — long blocks (and opt-in short/transient blocks), with alias reduction and frequency inversion
3. Rate/distortion quantization loop with a psychoacoustic masking model: scalefactors are amplified to push quantization noise under the Bark-band masking threshold (Schroeder spreading + ATH), with scalefac_scale and preflag
4. Huffman coding (the ISO tables + count1 quads)
5. Constant-bitrate frame assembly (header + side info + main data)

Scope & quality

• Now: encode mono / stereo / joint(M/S), CBR + VBR (Xing header), long blocks plus opt-in short/transient blocks (shortBlocks: true, mono AND stereo/joint); DECODE every block type (long/short/mixed/start/stop) — the codec round-trips in pure Dart and decodes any real-world MP3.
• Not yet: AAC/Opus. Follow the repo for progress.

Benchmark vs LAME / ffmpeg

Encode → decode (ffmpeg) → best-lag/gain-aligned SNR against the source, on three real 4 s / 44.1 kHz mono clips: a rendered instrumental band (drums+bass+chords+melody), macOS say speech, and an isolated bell attack. CBR, so file sizes are fixed by bitrate (ours runs a few hundred bytes smaller — no ID3/padding).

Reconstruction SNR (dB, higher = closer to source):

clip	kbps	ours	LAME	ffmpeg
band (drums+bass+chords+melody)	128	33.8	31.6	31.6
	192	42.5	34.5	34.5
speech	128	44.8	42.2	42.2
	192	55.8	53.3	53.3
bell attack	128	73.8	73.6	73.6
	192	78.3	77.0	77.0

Our encoder leads on raw SNR at every point — it spreads quantization noise more evenly. LAME/ffmpeg trade raw SNR for perceptual placement (noise pushed under the masking threshold), so a lower SNR there is partly by design; NMR (noise-to-mask) is the metric where the reference still edges ahead, and closing it is the next quality step.

Short/transient blocks help where pre-echo is actually audible — an isolated attack after quiet (the bell gains up to +0.9 dB in the worst 20 ms window). The conservative trigger deliberately leaves dense mixes on long blocks, and that's correct: measured across bitrates, forcing short blocks onto the band clip costs ~0.5 dB (the window-switching quantizer is not psy-shaped yet, so the extra short-block bits don't pay off), and inside a busy mix the ongoing signal forward-masks pre-echo anyway. So the real quality lever is psychoacoustic shaping of the short-block quantizer (short scalefactors), not a more eager detector — making the short blocks that already fire more bit-efficient, and closing the NMR gap generally.

API

• Uint8List mp3EncodeMono(Float64List pcm, {int sampleRate = 44100, int bitrate = 128}) — encode mono PCM in [-1, 1] to a CBR .mp3. sampleRate ∈ {44100, 48000, 32000}; bitrate is any MPEG-1 rate (32…320 kbps).
• mp3EncodeStereo / mp3EncodeJointStereo(left, right, ...) — stereo & mid/side.
• mp3EncodeMonoVbr / mp3EncodeStereoVbr(..., quality: 0..9) — variable bitrate.
• Mp3Pcm mp3Decode(Uint8List mp3) — decode back to interleaved float PCM.
• WAV helpers: readWavPcm16(bytes) → WavData, wavToMonoFloat(WavData) → Float64List.
• Frame constants: kMp3Bitrates, kMp3SampleRates, mp3FrameSize(...).

License

MIT. Ported from glint (MIT).