Resampler Modes for TUS Plugins
April 04, 2026
Beginning with the 2.2.2 versions of our plugins, we have implemented and provided some additional resampling options. These options can be found in the Settings Dialog under the DSP & Audio tab:
MAME HQ vs Legacy Resampler
Legacy (libresample)
- Kaiser-windowed sinc with linear coefficient interpolation
- 11-35 multiply-accumulates per sample depending on quality setting
- Kaiser window (Beta=6), 90% Nyquist cutoff
- ~70-100 dB stopband rejection
- Interpolates between filter coefficients at runtime (introduces small phase error)
MAME HQ (polyphase FIR)
- Pre-computed polyphase filter bank — no runtime interpolation
- Computes GCD of source/target sample rates to determine exact phase filters
- Up to ~400 MACs per sample, but all from pre-computed tables (cache-friendly)
- Hann-windowed sinc, cutoff at min(fs/2, ft/2)
- 100+ dB stopband rejection
- Fixed 5ms latency
- Per-phase normalization preserves DC gain exactly
What MAME HQ does better
- No interpolation error — every phase is pre-computed exactly, while legacy approximates between filter taps via linear interpolation
- Better anti-aliasing — 100+ dB stopband rejection vs ~70-100 dB, meaning less audible aliasing artifacts
- Exact phase filters — the polyphase decomposition means no phase distortion at any fractional delay
- Deterministic latency — fixed 5ms vs variable filter-dependent latency
Trade-off
MAME HQ uses more CPU (~400 MACs vs 11-35). The biggest audible difference would be on sample rate conversions with awkward ratios (e.g., 44.1kHz ↔ 48kHz) where the legacy resampler’s linear coefficient interpolation introduces subtle artifacts that the polyphase approach eliminates entirely.
MAME Lo-fi
There’s also a MAME Lofi mode (4-point cubic interpolation, ~4 MACs) for CPU-constrained scenarios that’s still better than linear interpolation but significantly cheaper than either HQ mode. The lo-fi resampler uses 4-point filtering with a box filter for downsampling, and has almost negligible latency and very low CPU usage.