Table of Contents · 12 Chapters
- 01Design Goals & Signal Boundaries
- 02PCM Resampling & Sample Rate Families
- 03Filter Bandwidth, Phase & Two-Stage Architectures
- 04Numerical Precision & The True Meaning of 280 dB
- 05DSD Passthrough, Decoding & Remodulation
- 06DSD Modulation & Stability
- 07Numerical Implementation of High-Rate 7th-Order
- 08CPU, GPU & Real-Time Processing
- 09Measurement Methodology & Empirical Evidence
- 10Output Formats & Prerequisites for Bit-Perfect Delivery
- 11Boundaries Between AI Metadata & Audio Processing
- 12References & Scope of Versions
Design Goals & Signal Boundaries
Preserving raw source data and active digital processing are two distinct playback philosophies, both deserving transparent presentation.
An audio player can deliver audio bit-perfectly without alteration, or intentionally process it. The former focuses on preserving identical sample bits; the latter focuses on the frequency response, noise profile, numerical stability, and real-time execution of the transformation. RLPlayer maintains dedicated paths for both paradigms, using an explicit signal chain to disclose exactly which processing stages are active during playback.
Upsampling reconstructs an existing signal onto a denser time grid. It cannot recover information that was never captured in the recording, nor does a higher output sample rate alone guarantee superior subjective sound quality. Its engineering value lies in delegating digital reconstruction filtering to transparent, verifiable, and user-selectable software algorithms, and aligning with the clock frequencies supported by the output hardware. How a DAC chip processes high sample rates or DSD internally remains governed by its silicon architecture and firmware.
The diagram illustrates the primary processing sequence; active playback bypasses any disabled stages. ReplayGain and peak headroom management operate on the input side; 20-band parametric EQ, per-channel routing, and room correction take effect at the source sample rate; volume control and quantization occur after resampling. Equalizers, delay lines, convolvers, and resamplers maintain state across audio blocks, avoiding discontinuities at block boundaries.
Making Processing Transparent
The user interface distinctly separates selected presets, requested output formats, and active execution status. GPU availability does not imply the current track actively uses GPU acceleration, nor does requesting 7th-order modulation guarantee the effective operational order is 7th-order. When encountering fallbacks, order reductions, or cross-family conversions, the UI discloses exactly what occurred rather than merely echoing user preferences.
PCM Resampling & Sample Rate Families
Going from 44.1 kHz to 705.6 kHz is an integer 16x ratio; going from 44.1 kHz to 384 kHz is a rational 1280/147 ratio.
Rational sample rate conversion reduces the ratio of output rate to input rate into irreducible integers L/M: interpolating the time grid by L, applying low-pass filtering to eliminate spectral imaging, and decimating by M. Polyphase implementations decompose filter coefficients into discrete phases, calculating only the sub-coefficients required for current output samples, without ever allocating zero-stuffed memory buffers.
44.1, 88.2, 176.4, 352.8, and 705.6 kHz belong to the standard 44.1 kHz clock family; 48, 96, 192, 384, and 768 kHz belong to the 48 kHz family. Staying within the same family yields smaller integer multipliers and simpler polyphase scheduling. Cross-family conversion is mathematically viable but requires rational resampling paths that alter sample values, forfeiting bit-perfect transmission.
| PCM Policy | Automatic Selection | Manually Specified Cross-Family Rate |
|---|---|---|
| Prefer Same Family (prefer) | Prioritizes device-supported same-family rates; prompts if no same-family rate is available | Allows explicit user override while explicitly flagging cross-family conversion |
| Strict Same Family (strict) | Permits only same-family conversion or bit-perfect passthrough | Rejects cross-family conversion |
| Compatible Conversion (compatible) | Allows cross-family fallback when no same-family rate is supported | Allowed, with active conversion explicitly flagged |
Swipe horizontally on smaller screens to view full table
Policy enforcement checks against the actual source sample rate and final hardware PCM clock, not merely the user's initial preference. Empty device capability lists, unknown sample rates, or unsupported manually targeted rates result in deterministic outcomes. Unknown clock families default to preserving source rates under auto mode; only compatible policies allow automatic conversion across clock families.
Preserving Source Sample Rate
When input and output sample rates match, the resampler operates in transparent passthrough. Full bit-perfect transmission across the entire pipeline still requires verifying channel layout, bit depth, volume unity, ReplayGain, EQ, convolution, and dither bypass. Losslessly zero-padding a 16-bit integer into a 24-bit container preserves numerical values, but transmitted sample words are no longer byte-for-byte identical to the file stream, warranting clear distinction from bit-depth passthrough.
Filter Bandwidth, Phase & Two-Stage Architectures
79 named presets provide distinct bandwidths, transition bands, time-domain distributions, and filter architectures. A preset's name alone does not constitute an empirical measurement.
A Finite Impulse Response (FIR) filter computes each output sample by weighting and summing adjacent input samples. The tap count is the number of weighting coefficients. Narrower transition bands and deeper stopband rejection necessitate longer filters, increasing coefficient generation time, memory footprint, computational load, and algorithmic latency.
From Budget Nomenclature to Actual Coefficients
Early implementations assigned a fixed coefficient budget across the entire prototype domain, causing per-phase taps to shrink as interpolation factor L grew. Current production pipelines dynamically derive per-phase filter lengths based on source-domain transition bands and design targets; explicit source-domain lengths, apodizing filters, and cascaded secondary stages follow specialized rules. When coefficient requirements exceed memory budgets, the pipeline explicitly rejects construction or switches to structured cascades—it never silently truncates filters or compromises quality to fit memory constraints.
Cutoff frequencies must not be conflated with passband edges. Standard resampling low-pass filters place the cutoff point at the center of the transition band; passband, transition band, and stopband are evaluated independently. A single reading at 20 kHz located within the roll-off zone cannot represent passband ripple.
| Phase Selection | Time-Domain Characteristics | Key Evaluation Criteria |
|---|---|---|
| Linear Phase (lp) | Symmetric FIR exhibits constant group delay across all active frequencies | Symmetry, latency, and pre/post-impulse energy distribution |
| Minimum Phase (mp) | Concentrates impulse energy toward the start of the response; group delay varies with frequency | Frequency response deviation, time-domain energy distribution, and actual latency |
| Intermediate Phase (ip) | Balances phase linearity and energy distribution trade-offs within bandwidth targets | Actual impulse envelope; ringing cannot be inferred from nomenclature alone |
Swipe horizontally on smaller screens to view full table
Current minimum and intermediate phase designs utilize factorized construction and high-precision convolution to prevent precision loss when transforming long prototypes with ultra-deep stopbands. While linear and minimum phase variants share similar frequency targets, claims of "identical amplitude response" or "zero pre-ringing energy in minimum phase" require rigorous measurement and cannot be asserted unconditionally for all presets. [1]
Two-Stage Upsampling Is Not Two Separate Sound Effects
In pure integer upsampling with sufficiently large ratios and cascading enabled, the first stage establishes source bandwidth while the second stage eliminates newly introduced images at higher sampling rates. The polyphase implementation enables cascading when M=1, L≥8, and valid factorization exists. 44.1→705.6 kHz supports cascading; 44.1→384 kHz has M=147 and does not run two stages merely because the preset has a "-2s" suffix. DSD workflows explicitly request compatible cascades rather than guessing based on high PCM rates.
Apodizing filters serve a distinct purpose: introducing smooth high-frequency roll-off prior to Nyquist. By altering high-frequency content, their performance must be benchmarked against their declared roll-off zones. They cannot be deemed defective under wide passband criteria, nor can stopband verification and end-to-end testing be skipped simply because a filter is apodizing.
Numerical Precision & The True Meaning of 280 dB
280 dB defines a specialized digital filter design target—it does not represent the analog dynamic range a DAC can achieve.
Stopband rejection is the attenuation of rejected frequencies relative to the passband. An amplitude ratio of a corresponds to 20 log10(a) dB; −280 dB corresponds to an amplitude ratio of approximately 10⁻¹⁴. While this number is meaningful in mathematical models and high-precision digital simulations, it cannot be translated into overall system SNR, effective bit depth (ENOB), or audible improvement.
Current verification sets a 280 dB rejection target for most primary filter families, and a 100 dB target for Gaussian families. These represent deliberate product quality targets for specific configurations, not universal mathematical limits of window functions. Whether a preset meets the target depends on source rate, target rate, phase response, cascade architecture, passband definition, and actual synthesized coefficients.
Design, Computation, and Measurement Represent Three Distinct Boundaries
Coefficient synthesis determines the ideal linear filter; numerical execution introduces coefficient quantization and accumulated rounding errors; measurement instruments possess finite resolution. Merely increasing FFT points does not automatically yield credible −280 dB readings. Deep stopband verification relies on actual exported coefficients, high-precision frequency response evaluation, endpoint analysis, and local refinement to prevent instrument noise floors from being mistaken for filter limits.
The production floating-point pipeline primarily utilizes IEEE 754 double (64-bit float), providing roughly 53 bits of significand precision. When computing minuscule residuals where large opposing terms cancel, rounding errors can exceed the magnitude of the result. Double-double (DD) arithmetic splits numbers into high and low components to enhance intermediate precision during synthesis and verification; this does not mean every playback sample runs in DD, nor does it imply 106-bit audio output at the sound card.
Qualifying integer PCM paths can employ fixed-point coefficients with 128-bit integer accumulation. While accumulation itself can be exact, quantizing coefficients into integer representations introduces rounding, as does converting to the final output format. This applies exclusively to integer grids accurately represented by the source without intervening DSP; one cannot quantize equalized or gain-adjusted floating-point samples back into integers and claim "zero error."
Strict Arithmetic Ordering
Aggressive fast-math compiler optimizations can reorder additions or fuse multiply-accumulates (FMA), altering significand bits. In linear filtering, this produces negligible variations; in 1-bit modulators with recursive feedback, minute deviations can cascade into divergent output bitstreams. Constraining reassociation, contraction, and reciprocal transformations—alongside scalar/SIMD parity checks, exception handling, and PRNG step verification—ensures reproducible computation. These constraints do not imply every compiler flag across all targets guarantees full IEEE strict mode.
DSD Passthrough, Decoding & Remodulation
DSD represents audio using extreme bitrates with 1-bit quantization; its noise distribution differs fundamentally from PCM.
DSD64 operates at 2.8224 MHz for the 44.1 kHz family, DSD128 at 5.6448 MHz, doubling up to DSD1024 at 45.1584 MHz; the 48 kHz family spans 3.072 MHz to 49.152 MHz. 1-bit quantization means the output switches between only two voltage levels; audio information is encoded in pulse density over time, rather than each individual bit representing an incremental amplitude step. [2]
| Task | Actual Processing Path | Data Integrity |
|---|---|---|
| DSD Same-Rate Passthrough | Read raw bitstream → DoP or Native DSD transmission | Preserves raw DSD bits when passthrough conditions are satisfied |
| DSD to PCM | Raw bitstream → Low-pass & decimation → PCM processing pipeline | Output is filtered PCM |
| PCM to DSD | PCM processing → Oversampling → Headroom management → SDM → 1-bit packing | Re-quantized bitstream; not raw source passthrough |
| DSD Rate Conversion | Raw bitstream → Intermediate PCM → Interpolation → Remodulation | Synthesizes a new DSD bitstream |
Swipe horizontally on smaller screens to view full table
Decoding & Decimation
Raw DSF/DFF bitstreams can be read in chunks by native decoders, then decimated into PCM via FIR, IIR, or CIC paths. Low-pass filtering suppresses high-frequency quantization noise, while decimation reduces sample rate. Each path exhibits distinct passband flatness, phase linearity, latency, and compute load; observing a 24-bit PCM container alone does not disclose the decimation chain's actual fidelity.
Same-family DSD rate conversion utilizes intermediate PCM remodulation. This intermediate PCM is an internal high-precision multi-level representation, not a downscaled CD-depth format. Directly comparing two DSD bitstreams cannot determine fidelity, as the same signal can yield distinct noise-shaped sequences. One must demodulate both streams into an identical bandwidth to evaluate amplitude accuracy, noise floors, and aliasing residuals.
The dedicated DSD rate planner supports integer-ratio conversion between standard same-family rates; cross-family requests are explicitly rejected. This limitation reflects current architectural choices rather than mathematical impossibility. Cross-family conversion should be treated as an explicit PCM conversion and remodulation workflow, never misrepresented as native DSD passthrough.
DSD Modulation & Stability
Absence of NaN or crash does not prove a modulator is functioning properly.
Modulator Selection & Valid Configuration
Sigma-Delta Modulators (SDM) use recursive feedback loops to convert high-precision multi-level signals into low-bit-depth streams. Noise shaping pushes quantization noise out of the audible band in exchange for low in-band noise; it does not eliminate noise. The Noise Transfer Function (NTF) governs how quantization error varies across frequency, while loop topology and state implementation dictate execution under finite-precision arithmetic. [2]
RLPlayer provides standard DSD, Chebyshev, EC, and AHM configurations. Modulator order, quantization levels, dither, and noise-shaping constraints cannot be evaluated in isolation from output formats. Higher order does not automatically confer superiority; one must verify in-band noise, out-of-band energy, loop stability, and computational throughput. The operational status must disclose effective order and active constraints.
In 3.0.10, the high-rate standard 7th-order NTF adopts a factorized cascade implementation, avoiding coefficient cancellation and precision loss caused by expanding high-order polynomials into direct feedback loops. High-rate requests no longer need to fall back to 5th order to circumvent this flaw; production guards continuously verify design validity. Factorized improvements must be substantiated by empirical in-band error tests, not extrapolated into universal sound quality claims.
AHM configurations with 6 to 8 bits provide 64, 128, or 256 quantization levels, utilizing adjacent step spacing; current code aligns requested bit depth with actual step count. When driving DoP and ASIO Native DSD paths, downstream hardware accepts exclusively 1-bit streams: AHM explicitly constrains to 2-level quantization under the same NTF and reports the reason. Multi-level hardware output is not currently exposed. Bit-packers reject invalid multi-level samples to prevent severe sign-only clipping distortion.
Upsampling, EQ gain boosts, convolution, and inter-sample peaks (ISP) can push calculated amplitudes beyond the maximum sample in the source file. High-order 1-bit loops can destabilize well before reaching numerical 0 dBFS. Full-scale PCM cannot be assumed to be a safe operational level for modulators without dedicated headroom management.
Current headroom policies scale operational peaks based on target DSD rates: 0.45 for rates below 4 MHz or unknown rates, and 0.62 for higher rates, with additional safety margins. These represent operating points derived from offline stability matrices, not universal mathematical theorems for arbitrary audio. Headroom attenuation is explicitly exposed in the signal chain.
Quantizer input limiting, feedback error clamping, non-finite value traps, and effective order guards together ensure loop protection. While NTF out-of-band gain and pole constraints filter out high-risk designs, linearized models cannot guarantee nonlinear modulator stability across all possible signals; running real audio loops remains essential.
| Verification Input | Key Observed Results |
|---|---|
| Steady-state sines, multi-tones, and silence | In-band noise, DC offset, overload increments, long-duration stability |
| Full-scale 0 dBFS and single-sample clipping | Interpretable clamping behavior, state variables remain bounded |
| Signal or silence following transients | Band-limited RMS recovery to reference, normal sign distribution, limiter count stops incrementing |
| Varying block lengths and channel configurations | Block-by-block output, PRNG state, and internal registers remain continuous and deterministic |
Swipe horizontally on smaller screens to view full table
A continuous stream of full-scale identical-sign bits or square waves can remain numerically bounded. Recovery criteria must therefore simultaneously evaluate band-limited energy, DC offset, sign tendencies, and limiter counters. Passing a test matrix verifies only the conditions contained within that matrix; it cannot substitute for the physical operating limits of amplifiers and speakers.
Numerical Implementation of High-Rate 7th-Order
Preserving the NTF factor structure prevents high-order expansions from losing numerical precision at ultra-high bitrates.
At DSD1024 bitrates, expanded polynomials for standard 7th-order Butterworth NTFs contain near-canceling coefficients. Feeding these directly into high-order difference equations amplifies finite-precision rounding errors. Increasing modulator order can paradoxically degrade in-band fidelity, even when output streams appear stable.
In 3.0.10, the modulator retains one 1st-order and three 2nd-order factors, updating states in a cascaded biquad structure rather than relying on rounded 7th-order polynomial coefficients for feedback. Each stage maintains unity feedthrough, with feedback derived strictly from prior states, avoiding extra sample delays. The theoretical design target remains standard 7th-order noise shaping; the improvement lies entirely in numerical implementation.
| Bitrate | Legacy 7th-Order (Diagnostic Forced) | Legacy 5th-Order (Safety Baseline) | 3.0.10 7th-Order |
|---|---|---|---|
| 45.1584 MHz | −69.0169 dBFS | −138.6755 dBFS | −166.8431 dBFS |
| 49.152 MHz | −63.4041 dBFS | −138.7398 dBFS | −167.3963 dBFS |
Swipe horizontally on smaller screens to view full table
Values in the table represent the error_all metric: subtracting the 1-bit output from the test tone in the time domain, then integrating total in-band error power. Unlike metrics that notch out the fundamental, this includes fundamental amplitude and phase residuals. 0 dBFS is defined as RMS=1, meaning a full-scale sine is −3.0103 dBFS; external data must align reference levels before comparison.
Test conditions: Standard DSD family, input peak 0.62, quantizer dither 0.0005 LSB, NTF cutoff 100 kHz, 50 ms pre-roll. Three full-cycle tones (~0.1, 1, 5 kHz) measured at each rate, evaluating three consecutive 1,048,576-point periodic Hann windows aligned to FFT bins. Error integrated from the first positive bin to 20 kHz (DC excluded, fundamental retained); power across the nine runs is averaged linearly and converted to dB.
The new 7th-order runs with production safety guards enabled, operating at an effective 7th order; steady-state windows exhibit zero overload, clamping, or non-finite increments. This confirms the factorized implementation resolves high-order numerical degradation under these conditions. It does not imply a universal error floor across all inputs, nor does it represent system dynamic range or subjective listening rank.
CPU, GPU & Real-Time Processing
The true value of hardware acceleration is increasing processing throughput and lowering system load while preserving bit-level algorithmic intent.
The CPU pipeline leverages SIMD vectorization to compute independent channels and frames concurrently. Fixed-point and floating-point paths serve distinct operational boundaries; one cannot narrow filter bandwidths, truncate taps, or alter quantization rules to inflate throughput figures while claiming acceleration of the same algorithm.
The GPU backend supports AMD HIP and NVIDIA CUDA, offloading parallel upsampling and convolution tasks. Transferring buffers to GPU memory, launching kernels, and synchronizing host threads carry fixed overheads, meaning short convolutions or small tasks may not run faster on GPU. Long impulse response convolution employs explicit length gating and persistent streaming sessions that maintain filter states and overlap tails across calls.
GPU acceleration does not mean every sample in an SDM recursive feedback loop runs in parallel on the GPU. Feedback loops possess strict sequential sample dependencies; real-world architectures offload pre-resampling to the GPU while keeping modulation and output packing on the CPU. Two-stage upsampling distinguishes whether one stage or both stages execute on the GPU.
Three Tiers of Parity Verification
Tier 1: Build & Interface—binary architecture and exported symbols exist. Tier 2: Execution—trivial test kernels read back expected constants, with verified execution logs during processing. Tier 3: Numerical Parity—identical input, coefficients, and state match reference outputs under strict tolerances. Driver initialization success or DLL load codes cover only the most basic tier.
Select configurations achieve bit-for-bit identity; other paths evaluate dimensioned errors, frequency response deviations, or end-to-end criteria. This whitepaper does not extrapolate local parity tests into claims that all CPU and GPU paths are permanently bit-identical. Empirical GPU execution in this environment was validated on AMD HIP; CUDA build and interface checks do not substitute for on-device NVIDIA execution verification.
Interpreting Processing Speed
Real-time multiplier is processed audio duration divided by elapsed processing time; values above 1.0 indicate the pipeline outpaces real-time playback. Reports must state channel count, sample rates, filters, modulators, batch sizes, hardware, and whether initialization and buffer transfers are included. Paired offline benchmarks in 3.0.8 across 14 paths and 120 runs confirmed statistically significant throughput gains on select paths; these results do not constitute universal real-time guarantees across arbitrary hardware.
Measurement Methodology & Empirical Evidence
Every reported metric must answer four questions: What was measured, where was it sampled, how was it computed, and what are the valid boundaries of the conclusion?
Filter verification exports actual coefficients directly from production objects rather than constructing surrogate filters to simulate the pipeline. Each measurement identity records input rate, output rate, preset, phase, and cascade configuration, alongside build hashes, ensuring every curve can be audited against the exact production code.
Frequency-domain evaluation isolates passband from stopband. Passband checks gain accuracy, ripple, and bandwidth; stopband checks the worst-case residual across specified bands rather than cherry-picking the deepest notch. Image rejection and aliasing attenuation are defined relative to conversion direction and bandwidth. Dense FFT grids, boundary checks, and local refinement enhance coverage, but finite-point testing does not constitute continuous mathematical proof.
Two-stage architectures cannot simply multiply two ideal frequency responses to claim implemented performance. Verification must audit production coefficients and real time-domain execution, factoring in rounding at each stage and using conservative envelopes or direct fallbacks. Phase and latency must be measured along the actual signal path to avoid mistaking slight time alignment offsets for algorithmic flaws.
| Archived Baseline | Observed Metrics | Supported Conclusions |
|---|---|---|
| 3.0.8 Filter Quality Matrix | 948 records; 863 applicable records met respective targets, 85 excluded | 720 met dense-grid targets, 143 met finite-grid processing envelopes; not a continuous spectrum proof |
| Production Report vs. Pipeline Cross-Check | 49,416 samples compared, 0 differing samples | Reported structures match actual pipeline output; cannot be extrapolated across all unmeasured configs |
Swipe horizontally on smaller screens to view full table
Exclusion is not a pass. It indicates a configuration does not apply to that criteria boundary and must state the specific reason. Gaussian filter targets differ from other families and cannot be aggregated into claims that "all filters achieve 280 dB." Figures reflect archived baselines; subsequent releases altering DSP algorithms must re-validate affected configurations rather than reusing legacy figures.
Music Analysis vs. Laboratory Test Signals
Real-world music contains complex transient material without a "perfect reference" for direct subtraction. Dynamics, bandwidth, clipping, effective bit depth, and spectral distributions provide valuable diagnostic insights, but statistical summaries alone cannot prove recording authenticity, mastering lineage, or harmonic distortion. THD+N measurements must state test signals, bandwidth, windowing, fundamental handling, and noise integration bands.
Comprehensive head-to-head comparisons against HQPlayer under identical hardware, input signals, bandwidths, output formats, and timing constraints have not been conducted across all paths. Algorithmic refinements and regression evidence in this project do not justify claims of universal superiority over HQPlayer.
Output Formats & Prerequisites for Bit-Perfect Delivery
Hardware capabilities dictate whether the final output stage can execute as requested.
Windows local output utilizes backends such as WASAPI or ASIO. WASAPI Shared routes audio through the Windows audio engine, which may alter sample rate, volume, or sample formats; Exclusive mode bypasses system mixing, but still requires verifying that the hardware natively accepts the target rate, bit depth, and channel count. Successfully opening Exclusive mode does not guarantee the upstream application pipeline applied no DSP.
ASIO Native DSD output in this project employs device-supported LSB1 or MSB1 bitstream sample types, matching the required bitrates. The ASIO specification defines additional NER8 extensions; this does not imply multi-level AHM hardware output is currently enabled. [4] DoP packs 16 consecutive DSD bits per channel into PCM-like frames with alternating 0x05 / 0xFA marker bytes; the carrier frame rate is DSD bitrate divided by 16 (DSD64 requires 176.4 kHz, DSD128 requires 352.8 kHz). DoP carriers are not PCM audio and must never undergo digital volume attenuation or resampling. [3]
Valid DoP markers confirm only the encapsulation format. Full verification requires proving post-unpacking bit identity, inter-block marker phase continuity, correct channel and bit ordering, and DAC DSD recognition. Same-rate passthrough also requires bypassing EQ, convolution, digital gain, and all other signal-altering stages.
| Signal Chain Status | Technical Meaning |
|---|---|
| Bit-Depth Passthrough | Source and output formats match with all processing bypassed; original sample bits preserved |
| Lossless Zero-Padding | Preserves exact numerical values while expanding bit depth; transmitted word formatting changes |
| Active DSP | Resampling, gain, EQ, convolution, dither, or remodulation modified the digital sample data |
| Shared / Network Output | Endpoint device may apply downstream mixing, buffering, or format conversion; requires hardware audit |
Swipe horizontally on smaller screens to view full table
UPnP casting transfers audio files to a network renderer, which cannot inherit bit-perfect assertions established for local exclusive endpoints. Network reachability, request success, renderer acceptance, and sustained streaming represent distinct verification tiers.
Reducing PCM bit depth introduces quantization error. TPDF dither adds low-level random noise to decorrelate quantization error from the audio signal; noise shaping redistributes this noise across the frequency spectrum. Because they intentionally modify samples, they forfeit bit-perfect passthrough. Even for 24-bit output, assuming dither is entirely harmless without measurement is invalid; target format alignment remains essential.
Boundaries Between AI Metadata & Audio Processing
AI aids in cataloging and understanding albums; deterministic DSP generates playback samples.
RLPlayer's metadata features search and organize album, artist, track, composition, and performance information, using local caching to reduce redundant network queries. It provides an interpretative layer: metadata can have aliases, release dates can conflict, and generated summaries can contain inaccuracies. Citations, provenance checks, and manual edit options are therefore more vital than polished prose.
AI models play no role in real-time execution of resampling kernels, modulator feedback loops, sample quantization, or output bit-packing. Audio samples are generated by deterministic algorithms and parameters; metadata requests execute asynchronously and will never stall the audio playback loop if remote services lag.
Metrics in audio analysis are computed by deterministic analyzers; AI can translate raw figures into human-readable summaries. Summaries must not fabricate unmeasured metrics or treat spectral roll-off clues as definitive proof of origin. Complex musical harmonics and unknown mastering processes require clear diagnostic boundaries.
When utilizing online AI or external metadata sources, required track titles, album names, or query contexts are sent to configured remote services in accordance with active settings. The application cannot be broadly characterized as having "zero network data transfer." Local playback and offline DSP behavior must be understood separately from user-enabled online features.
How to Interpret Evidence in This Document
This whitepaper documents implemented pipelines and archived measurement scopes. Topologies currently under research, broader device compatibility, and subjective auditory impressions are not represented as implemented capabilities. Software metrics help eliminate bugs and quantify trade-offs; sound quality remains for listeners to evaluate under controlled volumes, transparent settings, and fair comparisons.
References & Scope of Versions
The scope of this document encompasses the Windows digital audio processing pipeline in RLPlayer 3.0.10.
Software baseline: Version-controlled DSP, engine, and output implementations, including polyphase resampling, named filters, strict floating-point arithmetic, fixed-point convolution, SDM, DSD rate conversion, DoP, rate family policies, and signal chain telemetry. Data baseline: 3.0.8 quality and performance archives, 3.0.9 integration sign-offs, and incremental validation notes for this release. Historical curves do not represent updated algorithms.
The high-rate standard 7th-order numerical benchmark for 3.0.10 is detailed in Chapter 7. This benchmark was conducted using production parameters with independent spectral verification; AHM quantization step counts and 1-bit output constraints match current code. This paper provides no performance warranties for unexposed multi-level hardware outputs or untested devices.
- [1]Julius O. Smith IIIIntroduction to Digital Filters — Filters Preserving Phase ↗Symmetric impulse response and constant group delay of linear phase FIR filters.
- [2]Michael Clifford / Analog DevicesFundamental Principles Behind the Sigma Delta ADC Topology Part 1 ↗Principles of oversampling, feedback, and noise shaping; used for theoretical context, not as certification of RLPlayer performance.
- [3]DSD GuideDoP Open Standard ↗Encapsulation, marker structures, and carrier relationships of DSD over PCM (DoP).
- [4]Steinberg Media TechnologiesAudio Streaming Input Output Specification SDK 2.3 ↗Page 35, Sony DSD formats; public mirror of manufacturer specification distinguishing LSB1, MSB1, and NER8.
External theoretical reference audit date: November 1, 2026. RLPlayer is an open-source project licensed under GPL-3.0-or-later; third-party component and data licenses are disclosed in the release bundle. Project Developer: RainLain, Cats and Asteroids; Contributing Developer: Listen Storm; Media & Artwork: Gosed.