DAC beats Opus 16
At roughly half the storage, packed DAC has better SI-SDR and substantially lower spectral distance in both stress cohorts.
Reproducible audio codec benchmark · 2026
DACOPUSMP3
How much fidelity fits in an hour? A 170-clip study of storage, perceptual degradation, spectral error, intelligibility and neural decoding speed.
44.1 kHz · MONO · ALL CODEBOOKS
Packed DAC bitrate7.76kbps measured
Storage per audio hour3.49MB / hour
L40S decode speed124×realtime median
DAC decoder network54.1Mparameters
01 · Storage economicsSmall tokens,
Small tokens,
large differences.
Packed all-codebook DAC is the smallest representation in this study. Opus 16 kbps uses 2.16× its storage; MP3 128 kbps uses 16.61×. Costs below use marginal R2 Standard storage before free tier or billing rounding.
| Representation | kbps | MB/hour | $/hour-month |
|---|---|---|---|
| DAC packed tokens | 7.76 | 3.49 | $0.000052 |
| DAC uint16 tokens | 12.41 | 5.58 | $0.000084 |
| DAC native file | 12.83 | 5.77 | $0.000087 |
| Opus 16 CBR | 16.75 | 7.54 | $0.000113 |
| Opus 48 CBR | 48.80 | 21.96 | $0.000329 |
| Opus 64 CBR | 64.82 | 29.17 | $0.000438 |
| MP3 128 CBR | 128.82 | 57.97 | $0.000870 |
02 · Objective qualityOne score is
One score is
never enough.
Twenty naturally challenging LibriSpeech utterances and twenty matched speech-plus-music mixtures expose different failure modes. We report waveform fidelity, intelligibility, multi-resolution spectral distance and a learned MOS estimate separately.
Intelligibility holds
STOI stays high for every codec. Higher-rate Opus and MP3 approach one, while DAC remains strong at its much smaller payload.
MOS is secondary
SQUIM is trained for speech and can behave oddly on music mixtures. Near-zero or negative deltas are model-level ties, not proof of improvement.
03 · Neural decodingRealtime,
Realtime,
comfortably.
The 54.104M-parameter synthesis decoder runs at 2.68× realtime on this 12-thread CPU and 124.38× on an NVIDIA L40S. Including quantizer lookup, the complete decode path is 54.344M parameters.
Benchmark boundary
What the timer includes
Codebook reconstruction, neural synthesis, GPU-to-CPU output materialization, loudness normalization and output-length restoration. Disk I/O, model loading and encoding are excluded.
CPU: 2.65–2.74× realtime
GPU: 111.0–139.6× realtime
Full codec: 76.652M parameters
04 · MeasurementMetrics with
Metrics with
clear jobs.
Codec evaluation is multi-dimensional. The suite follows common neural codec practice while documenting where a metric is uncalibrated, speech-specific or unavailable in a safe reproducible build.
SI-SDR ↑
Strict, scale-invariant waveform reconstruction fidelity. Phase and timing changes count even when they sound benign.
STOI ↑
Short-time speech intelligibility. Applied only as a speech-component diagnostic.
Log-STFT L1 ↓
Three spectral resolutions capture both envelope and fine-structure mismatch.
SQUIM MOS loss ↓
Learned speech-quality prediction, paired against each pristine reference.
ViSQOL
Common full-reference codec metric. Not scored here because no safe official wheel supports this Python build.
MUSHRA
A controlled human listening study remains the gold standard for final perceptual decisions.
20 examples · 6 conditions each
Open listening lab →Hear the tradeoffs.
Switch between the reference, DAC, three Opus rates and MP3 while keeping playback position aligned.
05 · ProvenanceBuilt to be
Built to be
reproduced.
LibriSpeech test-other was checksum-verified and sampled deterministically across 20 speakers. It is naturally challenging read speech—not artificially noisy speech. MTG-Jamendo tracks come from an artist-disjoint test split with adaptation-compatible licenses, mixed at +5 dB active-speech SNR.
All exact hashes, offsets, gains, timings and raw measurements are downloadable below.