LILAC: An Idempotent Neural Speech Codec

Audio demonstrations

Companion page for "LILAC: An Idempotent Neural Speech Codec" (arXiv:2608.05727 · code · checkpoint)

Mechanism, in one paragraph. LILAC encodes speech through an invertible (conv-only, no attention) analysis chart to a 0.75 kb/s discrete code at 9.375 Hz frame rate. The decoder is constructed as the exact inverse of the encoder, so encode(decode(codes)) == codes holds by construction, not by training incentive. This page shows two things: (1) how LILAC's reconstructions at 0.75 kb/s compare to several published neural codecs on a single decode pass, and (2) what happens under repeated decode→re-encode cycling (K = 1, 10, 100) — LILAC's waveform output is byte-identical across all three, verified below via sha256 of the decoded PCM16 audio.

Codec comparison

CodecBitrateNative sample rateNotes
LILAC (ours)0.75 kbps24,000 Hzconv-only lifting chart + invertible anti-imaging stem; idempotent by construction
EnCodec1.5 kbps24,000 Hzfacebook/encodec_24khz, bandwidth=1.5 kbps
DAC24 kbps24,000 Hzdescript/dac 24kHz, full 32-codebook RVQ (near-lossless reference point, not rate-matched)
Mimi1.1 kbps24,000 Hzkyutai/mimi, 8 quantizers
FocalCodec0.65 kbps16,000 Hzlucadellalib/focalcodec, 50 Hz variant

Per-clip demonstrations (8 clips)

Each clip: ground truth, then every codec's decode at K = 1 (single pass), K = 10, and K = 100 repeated decode→re-encode cycles. The LILAC row's hash line shows the sha256 (first 16 hex chars) of each cycle's PCM16 bytes.

Clip 2300-131720-0034 LibriSpeech test-clean · speaker 2300 (M) · 8.6s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: 59a2282a340d9a77…  ·  K=10: 59a2282a340d9a77…  ·  K=100: 59a2282a340d9a77… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram

Clip 1320-122612-0013 LibriSpeech test-clean · speaker 1320 (M) · 6.5s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: a4f9b0cdc7357f34…  ·  K=10: a4f9b0cdc7357f34…  ·  K=100: a4f9b0cdc7357f34… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram

Clip 8224-274384-0007 LibriSpeech test-clean · speaker 8224 (M) · 6.2s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: 23ed00828ff220cd…  ·  K=10: 23ed00828ff220cd…  ·  K=100: 23ed00828ff220cd… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram

Clip 7729-102255-0025 LibriSpeech test-clean · speaker 7729 (M) · 5.5s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: d1ac20b6ee525ef2…  ·  K=10: d1ac20b6ee525ef2…  ·  K=100: d1ac20b6ee525ef2… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram

Clip 2830-3980-0050 LibriSpeech test-clean · speaker 2830 (M) · 7.5s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: 7c05265e6c5b17a2…  ·  K=10: 7c05265e6c5b17a2…  ·  K=100: 7c05265e6c5b17a2… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram

Clip 3570-5695-0002 LibriSpeech test-clean · speaker 3570 (F) · 7.8s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: 88dada03d88763f6…  ·  K=10: 88dada03d88763f6…  ·  K=100: 88dada03d88763f6… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram

Clip 3575-170457-0007 LibriSpeech test-clean · speaker 3575 (F) · 7.8s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: 92a496dfe244cbdc…  ·  K=10: 92a496dfe244cbdc…  ·  K=100: 92a496dfe244cbdc… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram

Clip 908-31957-0005 LibriSpeech test-clean · speaker 908 (M) · 4.5s

Ground truth ground-truth log-mel spectrogram

CodecK = 1K = 10K = 100
LILAC (ours)
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
sha256 K=1: da5d4af9f2b25ce8…  ·  K=10: da5d4af9f2b25ce8…  ·  K=100: da5d4af9f2b25ce8… byte-identical
EnCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
DAC
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
Mimi
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram
FocalCodec
log-mel spectrogram
log-mel spectrogram
log-mel spectrogram