Why Hi-Res Streams Can Sound Different From One Another

"The bits are honest. The question is always which bits, processed how, and by whom, before they reach you."

The previous article established something reassuring: when Qobuz delivers a 24-bit/96kHz FLAC file to your DAC, the bits that arrive are identical to the bits that left the studio master. Lossless is lossless. The mathematics are beyond dispute.

And yet experienced listeners consistently report that the same album sounds different across streaming platforms, or even across different software playing the same file from the same service. Some dismiss this as expectation bias, as the audiophile's tendency to hear what he wants to hear. The dismissal is too easy. There are several technically legitimate points in the chain where audible differences can and do arise, none of which contradict the lossless guarantee, because most of them operate either before the bits leave the platform or after the bits leave the file.

Understanding where these differences come from requires a more granular look at what "same recording" actually means in practice.

Before proceeding, one distinction is essential. If two playback chains can be demonstrated to deliver identical PCM data to the DAC, at the same playback level, without DSP, sample rate conversion, or format conversion, established digital audio theory predicts no audible difference attributable solely to the streaming platform itself. The differences discussed below arise from the many places where real-world playback chains depart from that ideal condition. This does not invalidate the listener's observations. It simply focuses attention on where the causes are most likely to be found.

1. The Master Is Not Always the Same Master

The most significant source of sonic differences between streaming platforms has nothing to do with the streaming process itself. It precedes it entirely. Record labels do not always supply every platform with the same master file.

This is not conspiracy; it is logistics and economics. When Apple Music launched its lossless tier and subsequently its Spatial Audio catalogue, many labels commissioned new mixes and masters specifically for the format. When Tidal introduced MQA as its flagship format, labels supplied MQA-encoded files that had been processed through MQA's proprietary encoding chain, a chain that its critics argue introduces its own coloration regardless of the lossless claim. Qobuz, which has consistently refused MQA and positions itself around unprocessed studio masters, may hold a genuinely different master for the same title than a competing service.

Mastering engineers are frequently asked to produce what the industry calls a "streaming master", optimized for loudness normalization targets, which differs from the CD master and often from the vinyl lacquer cut. The same album can therefore exist in several legitimate versions simultaneously, each technically flawless, each audibly distinct. The streaming service is delivering exactly what it was given, and what it was given is not the same object.

This is not a flaw in the system. It is the industry's normal operating reality, and it is the primary reason why platform comparisons are so difficult to conduct fairly. You are often not comparing delivery mechanisms. You are comparing different source materials.

2. Loudness Normalization and Its Hidden Costs

Every major streaming platform normalizes playback loudness, which means the platform adjusts the output level of each track so that all tracks play back at approximately the same perceived loudness. Apple Music targets -16 LUFS. Spotify targets -14 LUFS. Tidal and Qobuz both offer normalization as an option that most users leave enabled. YouTube Music normalizes to approximately -14 LUFS.

When a track has been mastered louder than the target, normalization applies gain reduction: the signal level is turned down. This is acoustically benign, equivalent to turning your volume control.

But when a track has been mastered quieter than the target, some implementations apply positive gain. In modern 24-bit playback chains this normally has negligible consequences, but it reduces available headroom and can, in extreme cases, clip transients that were not clipped in the original file.

A related phenomenon is the inter-sample peak. A file may contain no clipped samples and still exceed digital full scale after waveform reconstruction in the DAC. Some playback systems reserve a small amount of headroom specifically to avoid this condition. Others do not. This is another reason why apparently identical files can behave differently in real playback chains despite containing the same underlying audio data.

More importantly, normalization algorithms are not all the same. The psychoacoustic models used to calculate integrated loudness, the look-ahead windows, and the response to dynamic variation across a long track differ by platform and by application version. Two services applying normalization to the same file will not produce identical results at the output.

The differences are generally subtle, but they are measurable and can manifest as small variations in perceived weight, impact, and dynamic behaviour.

The cleanest solution for a serious listener is to disable loudness normalization entirely and manage playback level through the volume control of the preamplifier. This is one of several reasons why the passive or transformer-based volume stage, working on an already bit-perfect signal, remains a philosophically coherent choice in a modern digital system. The signal arrives intact; only its level needs adjustment. The role of the preamplifier as gatekeeper rather than generator is explored separately.

3. Sample Rate Conversion and the Algorithm Question

A 24-bit/96kHz file contains 96,000 samples per second. If your DAC is configured to accept only 44.1kHz or 48kHz, or if the operating system's audio engine is locked to a fixed output rate, something must convert the sample rate before the signal reaches the converter. That something is an algorithm, and algorithms are not equal.

Sample rate conversion is mathematically demanding. Converting from 96kHz to 48kHz is a straightforward factor-of-two relationship and can be done with minimal consequence if the filter implementation is competent. Converting from 44.1kHz to 48kHz, or from 88.2kHz to 48kHz, involves more complex resampling ratios that require careful implementation to avoid introducing aliasing, ringing, or passband ripple.

The location and magnitude of these artifacts depend on the design choices made in the SRC algorithm.

It is important not to overstate the problem. Modern sample rate converters from Apple, Microsoft, Roon, HQPlayer, SoX, and professional audio software generally reduce such artifacts to levels that are extremely low and often below audibility. The engineering question today is less whether conversion is possible and more how transparently it is implemented.

Windows, macOS, iOS, and Android each contain their own system-level sample rate converters of varying quality. Applications can bypass these entirely by claiming exclusive access to the audio hardware, which is what WASAPI Exclusive mode does on Windows and what ASIO drivers provide. When a streaming application does not request exclusive mode, the operating system's shared audio mixer inserts itself into the chain and applies its own sample rate conversion whenever the file's native rate differs from the mixer's configured rate.

This is a genuine, measurable source of sonic difference between otherwise identical setups. In most browser-based playback scenarios on Windows, a user playing Qobuz with the system mixer set to 48kHz and a file playing at 44.1kHz will not be hearing bit-perfect audio despite using a lossless service. The operating system mixer is converting the signal before it reaches the DAC.

The same user playing Qobuz through the dedicated desktop application, configured for WASAPI Exclusive mode, can pass the native sample rate directly to the DAC without resampling.

This is why the choice of playback software matters on resolving systems, independently of the source quality. Roon's signal path indicator makes this visible, displaying a color-coded signal path that shows whether playback remains bit-perfect or whether DSP, format conversion, or other processing has been introduced. It is one of the genuinely useful features of that platform from an engineering-transparency standpoint, and the subject of a separate guide on using Roon properly.

4. Spatial Audio, Atmos, and Algorithmic Processing

Apple Music's Spatial Audio and Dolby Atmos streams are a distinct category that deserves separate treatment because they are explicitly not bit-perfect delivery of studio masters in the traditional sense. They are processed renders.

A Dolby Atmos mix exists as a collection of audio objects accompanied by positional metadata. When Apple Music delivers Atmos content to headphones, the system applies a head-related transfer function, a complex set of filters derived from acoustic measurements of how sound interacts with the human head and outer ears, to create the impression of three-dimensional space.

This processing is applied in real time, varies by implementation, and is not present in the stereo mix of the same recording.

The result can be extraordinary or disorienting depending on the recording, the implementation, and the listener's anatomy. HRTF processing is based largely on population averages, which is why spatial rendering sounds convincing to some listeners and artificial to others.

Apple's Personalized Spatial Audio attempts to improve localization by building a more individualized HRTF model from scans of the listener's ears and head geometry, though the sophistication of this personalization remains limited compared with a fully individualized acoustic measurement.

The point is that when a listener compares Apple Music's Atmos version of an album to Qobuz's stereo master, they are comparing two fundamentally different artistic objects. The Atmos version has been processed.

This is not inferior delivery. In many cases it represents additional creative work by the mixing engineer. But it is categorically different from lossless delivery of a stereo master, and treating it as a simple quality comparison misunderstands what is actually being compared.

5. The Playback Application as a Signal Processor

Streaming applications are not passive conduits. They are software, and software can do whatever its developers choose before handing the signal to the operating system.

Most reputable hi-res streaming applications, including Qobuz's desktop client and dedicated network-streamer implementations, pass the decoded PCM stream to the audio output without modification. This is the correct behavior and corresponds to the lossless guarantee.

But not all applications are designed this way, and even applications that are designed this way may apply processing under specific conditions.

Some applications apply their own equalization or spatial enhancement features that are enabled by default and require the user to explicitly disable them. Some apply crossfeed algorithms designed to reduce headphone listening fatigue. Some insert soft limiting on transients to protect against inter-sample overs. Others may alter output behavior to accommodate device compatibility or battery-management requirements.

None of this is fraudulent, and in many cases it is benign or even beneficial for the casual listener.

But for the serious listener building a system around signal integrity, it matters that the application is examined rather than trusted by default. The assumption that a lossless source automatically implies a transparent application is not always warranted.

6. The Hardware Streamer and Its Software Layer

The discussion so far has addressed what happens at the streaming-platform end and within the playback application. But for many serious listeners, the signal does not flow directly from a general-purpose computer to a DAC. It flows through a dedicated hardware streamer, a purpose-built device that sits in the rack alongside the preamplifier and the DAC, receives audio over the network, and outputs a digital or analogue signal.

These devices are not passive receptacles. Each one runs a software layer of its own design, and the decisions made in that layer can be consequential whenever they affect DSP, sample rate conversion, filtering, volume management, clocking architecture, or output implementation.

The clearest way to understand the landscape is to see it as a spectrum between two poles. At one pole sits the streamer that passes the decoded PCM stream directly to the output with no intervention: a digital transport in the fullest sense. At the other pole sits the streamer that treats the incoming signal as raw material to be optimized, upsampled, filtered, and shaped before release.

Both approaches have coherent rationales. Neither is inherently wrong. But the distinction matters enormously for the listener who believes the value lies in what the studio put into the recording rather than what the playback chain adds afterwards.

Roon Nucleus and ROCK

Roon occupies a category of its own because it is software first and hardware second. The Roon Nucleus is a purpose-built server running Roon OS, a custom Linux-based operating system stripped of everything unrelated to running the Roon Core, and ROCK is its do-it-yourself equivalent for Intel NUC hardware.

What makes Roon architecturally distinct is its explicit transparency about what the signal is doing at any moment. The MUSE audio engine supports bit-perfect playback while also offering an extensive collection of DSP tools. The signal path display makes the state of the chain visible in real time, showing whether playback remains bit-perfect or whether DSP, format conversion, or other processing has been introduced.

The risk with Roon is not its defaults but its options. The platform includes parametric equalization, upsampling, room correction, crossfeed, and convolution filter support, any of which can be enabled accidentally or in the belief that more processing is automatically better processing.

A Roon installation with everything switched off is among the most transparent playback chains available. The same installation with upsampling to DSD512, convolution filtering, loudness leveling, and room correction enabled becomes something else entirely. Roon does not make these choices for the listener. It merely makes them possible. That is a form of respect, but it requires the listener to understand what each switch does before reaching for it, a discipline closely related to treating DSP as the last resort.

Aurender

Aurender's approach is among the most explicitly proprietary in the high-end streamer market. The company runs a custom Linux-based operating system tuned specifically for audio playback and controlled through its Conductor application.

For many years Aurender deliberately chose not to support Roon's RAAT protocol, preferring complete control over the playback environment. The philosophical position has always been that the performance gains from a vertically integrated, single-purpose software stack outweigh the convenience of participating in a broader ecosystem.

Whether that claim is audibly substantiated is a matter for the listener rather than the specification sheet. What is certain is that Aurender owners live entirely within the Conductor environment, and the quality of that environment, its reliability, usability, streaming-service integration, and library management, becomes part of the overall listening experience.

The engineering lesson is straightforward: software architecture is not necessarily audible in itself, but any architecture that changes processing behavior, output implementation, buffering strategy, or clock management can influence the final result.

Innuos

Innuos, the Portuguese company that has become one of the more thoughtful voices in the streamer market, takes a similarly proprietary position but with a different emphasis.

Its Sense operating system and accompanying Sense application represent a holistic approach to audio playback, applying system-wide optimization throughout the software stack and utilizing a heavily customized player engine targeting bit-perfect playback with zero EQ. The company's own public statements consistently emphasize signal integrity, low-noise operation, and minimizing unnecessary processing between source and output.

Innuos also supports Roon, which creates an interesting internal comparison: the same hardware, the same source material, played back through either the proprietary Sense engine or Roon's MUSE environment. Many Innuos owners report a preference for Sense. The origin of that preference remains difficult to isolate. It may reflect differences in software architecture, output behavior, processing load, user interaction, or simply personal preference.

To Innuos' credit, the company generally avoids making grandiose technical claims that attempt to explain every reported difference. That restraint is refreshing in a market that often substitutes certainty for evidence.

Auralic

Auralic runs its own Lightning OS, a software platform designed specifically around the company's Tesla hardware architecture and controlled through the Lightning DS application.

The important distinction with Auralic is that Lightning DS functions primarily as a control layer while the actual audio handling occurs within Lightning OS itself. The control application does not sit directly in the audio path, which is architecturally cleaner than systems where playback and control functions are more tightly coupled.

Lightning OS includes optional resampling, parametric equalization, loudspeaker-placement correction, and other DSP tools. None of these are active by default.

Auralic's upsampling implementation, which can bring PCM material up to 384kHz before output, is among the more sophisticated examples found in consumer streamers. Some listeners find the result transformative. Others prefer native-rate playback. Neither position is objectively correct.

The signal path from native-rate PCM to the DAC without upsampling remains bit-perfect. Once upsampling is enabled, bit-perfect delivery ceases by definition and the audible character of the chosen algorithm becomes part of the playback chain.

Lumin

Lumin occupies an interesting position in the market. Rather than creating an entirely proprietary transport architecture, the company built its platform around OpenHome, an extension of UPnP originally developed by Linn to improve audio-streaming functionality and reliability. This gives Lumin products broad interoperability while still allowing the company to maintain control over the playback engine and user experience.

Lumin has also become something of a platform provider. Manufacturers including Luxman, Esoteric, TEAC, and others have licensed elements of the Lumin streaming architecture for use in their own products. This matters because it reminds us that software platforms often extend far beyond the badge on the front panel. A listener comparing two apparently different products may in reality be comparing hardware implementations built upon very similar software foundations.

Many of the implementations derived from the Lumin platform include optional sample-rate conversion and upsampling features. Whether these improve the final presentation depends on the quality of the algorithms involved and on the listener's priorities. For some listeners, preserving the source unchanged is the goal. For others, carefully executed interpolation is itself part of the optimization process.

Naim

Naim's streaming platform, currently represented by the NP800 architecture, is among the most fully integrated designs in contemporary audio. The NP800 is not merely a network interface attached to a DAC. It is a complete digital-audio system in which the streaming stage, DSP implementation, clocking architecture, filtering strategy, and analogue output stage are designed together.

Naim employs proprietary digital filtering and extensive DSP processing before the signal reaches the converter stage. The company's engineering philosophy is that careful control over these processes produces superior results compared with relying solely on the standard implementations supplied by DAC-chip manufacturers.

Whether one agrees with that philosophy is ultimately secondary. What matters is recognizing that Naim does not position itself as a pure transport architecture. It is an integrated signal-processing architecture. The resulting sound is therefore the product of both the source material and Naim's chosen implementation.

Cambridge Audio

Cambridge Audio's StreamMagic platform is developed internally rather than licensed from a third party. The company's decision to create its own streaming architecture reflects a desire to control the complete playback chain while maintaining compatibility with services such as Roon, Spotify Connect, Tidal Connect, and Qobuz.

This creates an interesting situation. The same hardware may receive identical music through several different pathways, each using different control protocols and software layers. Experienced listeners sometimes report hearing differences between these pathways. Whether such differences remain audible when the resulting PCM stream is verified as identical has not been demonstrated consistently in controlled testing.

Nevertheless, the observation serves as a useful reminder that assumptions about transparency should be verified rather than presumed. The engineering question is not whether a pathway sounds different. The engineering question is whether something within that pathway is measurably different.

Bluesound and NAD

Bluesound, part of the Lenbrook group alongside NAD, approaches the problem from a different direction. BluOS is designed primarily as a whole-home audio ecosystem. The emphasis is on reliability, ease of use, multi-room synchronization, and broad compatibility rather than on absolute minimalism of the signal path.

Bluesound products are capable of bit-perfect playback when configured appropriately. However, some models include optional processing, upsampling, or timing-management features that may alter the signal before output. These are deliberate engineering choices rather than defects.

The listener who compares a Bluesound configured for processing against a minimalist transport configured for strict passthrough is not comparing identical signal paths. The comparison may still be useful, but it is important to understand what is actually being compared.

The Pattern

What emerges from surveying these platforms is that the streamer market has fragmented into several distinct philosophies. There are the purist transports, whose primary objective is to deliver the incoming bit stream to the DAC with minimal intervention. There are integrated signal processors, whose designers believe carefully chosen DSP, filtering, upsampling, or clock-management strategies improve the final result. And there are ecosystem platforms, which prioritize reliability, usability, and multi-room functionality over strict signal-path minimalism.

None of these approaches is inherently wrong. But they are not the same. The listener who assumes that every hardware streamer delivers identical data to a DAC in exactly the same way is making an assumption that the engineering of these products does not always support. Understanding which philosophy your hardware embodies is not an exotic concern reserved for engineers. It is part of understanding what you are actually listening to.

7. The DAC Still Does the Final Work

Everything discussed so far concerns what happens before the signal reaches the DAC. Yet regardless of whether the source is a CD transport, a local FLAC file, Qobuz, Tidal, Apple Music, or a dedicated streamer, the DAC ultimately reconstructs the analogue waveform. That reconstruction requires filtering, and the design of that filter remains one of the last places where engineering choices can measurably influence the final result.

Modern DACs employ a variety of reconstruction filters: linear-phase, minimum-phase, apodizing, hybrid, and proprietary implementations. These approaches differ in impulse response, phase behavior, and ultrasonic filtering characteristics. The source file remains unchanged, but the mathematical process used to reconstruct the analogue waveform does not.

In practice, modern DACs generally achieve excellent technical performance, and differences between competent designs are often subtle. Nevertheless, reconstruction filtering remains a more plausible source of audible variation than the idea that identical PCM streams somehow acquire different sonic signatures in transit. Once the data reaches the DAC unchanged, the DAC itself becomes one of the dominant variables in the final presentation. The way clocking and conversion sit at the top of that chain is the subject of The Digital Hierarchy.

This distinction matters because discussions about streaming quality often focus on what happens before the DAC while overlooking the fact that every digital playback chain ultimately converges at the reconstruction stage. A perfectly preserved signal can still be presented differently by different converters, not because the information changed, but because the method of reconstructing the analogue waveform differs.

8. Why the CD Keeps Winning

Given everything above, the CD's persistent reputation among serious listeners begins to look less like nostalgia and more like a rational response to a specific set of variables.

A compact disc is a physical object that encodes one particular master at the moment of pressing. That master was prepared for that release, for that format, and it does not change after the disc leaves the factory. There is no server deciding which version to stream, no platform-specific master, no loudness normalization applied without the listener's knowledge, and no software update that quietly alters the playback chain. The disc is what it is, and it is the same tomorrow as it was today.

The transport's primary task is to recover and deliver a 16-bit/44.1kHz PCM stream using the error-correction and buffering mechanisms built into the CD standard. Reed-Solomon error correction allows a CD player to recover data accurately even when a disc contains scratches, fingerprints, or manufacturing imperfections. When the error-correction system remains within its design limits, the recovered PCM stream is identical to the data originally pressed onto the disc. The process is largely invisible to the listener, which is precisely the point.

There is no operating system audio mixer inserting itself between the data and the converter, no application-level DSP running in the background, and no exclusive mode that must be configured correctly to bypass a resampling stage. The signal path from disc to DAC is structurally simpler than any streaming chain, and simplicity is a form of transparency.

The 16-bit/44.1kHz format itself is also more capable in practice than its theoretical limitations suggest. The Nyquist-Shannon sampling theorem guarantees that 44.1kHz captures everything up to 22.05kHz without aliasing, comfortably exceeding the upper boundary of human hearing. The approximately 96dB dynamic range of 16-bit audio exceeds the dynamic range of virtually every commercial music release, and the noise floor of a competent DAC sits well below the practical limits of the format.

The format's ceiling is not the ceiling listeners run into. What they run into is the mastering.

This is the core of the CD argument: the master pressed onto a CD in 1989, 1994, or 2003 was frequently produced before the loudness war reached its worst phase, before streaming normalization targets began influencing mastering decisions, and before labels started producing multiple variants of the same title for different distribution channels.

Many of those older masters possess more dynamic range, more headroom, and more low-level detail than the same album's current streaming version, regardless of whether that streaming version is offered at 24-bit resolution.

A 24-bit/96kHz stream of a dynamically compressed, aggressively mastered recording does not outperform a 16-bit/44.1kHz CD of a well-mastered one. Resolution is not the variable that matters most. The master is.

And for a significant portion of the recorded catalogue, the master that lives on the original CD pressing is simply the best version available. Not because the format itself is superior in absolute terms, but because it was pressed from a source that no longer exists in that form anywhere in the streaming ecosystem. That the recording and its master set the ceiling for everything downstream is the theme of Beyond Digital Purity.

There is also a practical dimension that experienced listeners seldom discuss but often rely upon. Playing a CD requires engagement. You select a disc, place it in the transport, and listen to an album as a deliberate experience. The format encourages attention in a way that streaming, with its infinite catalogue and algorithmic recommendations, subtly discourages. Whether this affects perceived sound quality is a question for psychology rather than engineering, but it is not irrelevant to the listening experience as a whole.

9. What Listeners and Engineers Actually Find

The question of which platform comes closest to the original file has been examined in several ways, and the results are more instructive than they are conclusive.

At the measurement level, the answer is straightforward. When two services deliver the same master and both operate in genuinely bit-perfect mode, the resulting PCM streams are mathematically identical. Multiple independent comparisons using software such as DeltaWave have demonstrated this repeatedly. The lossless promise, in that narrow and specific sense, holds.

The listening tests produce a more complicated picture, and the complications are revealing. Controlled comparisons often produce inconsistent rankings among experienced listeners. One listener prefers Qobuz. Another prefers Tidal. A third prefers Apple Music. The rankings shift depending on the material, the playback chain, and the versions of the recordings being compared.

What matters is not that listeners disagree. What matters is why. When investigators examine these comparisons closely, the explanation frequently turns out to be differences in mastering rather than differences in delivery. A listener believes he is comparing platforms. In reality, he may be comparing different source material.

This distinction explains why platform comparisons so often generate passionate disagreement while remaining frustratingly difficult to reproduce under controlled conditions. The variable under examination is rarely isolated as cleanly as participants assume. The wider gap between what is heard and what is measured is examined in What You Hear, What You Measure.

Editorial listening impressions from reviewers and enthusiasts tend to converge around Qobuz and Tidal as the most serious contenders for listeners pursuing maximum fidelity. Qobuz's reputation among audiophiles rests largely on its emphasis on unmodified studio masters and its historical refusal to adopt MQA. Tidal's catalogue, meanwhile, has undergone significant evolution as the company moved away from MQA and toward standard FLAC delivery.

The honest conclusion is that between modern streaming services operating in genuine bit-perfect mode, with loudness normalization disabled and the playback chain properly configured, the differences become very small. What remains is usually not the platform itself. It is the master.

Where This Leaves the Listener

The picture that emerges is not chaotic. It is structured. The lossless claim made by Qobuz, Tidal, Apple Music, Amazon Music, and other major services is technically honest. The file delivered to the listener is a bit-perfect representation of the master held by the platform.

The audible differences arise elsewhere. The master files themselves may differ between platforms. Loudness normalization systems differ. Sample rate conversion quality differs. Playback applications differ. Hardware streamers differ. DAC implementations differ. Spatial audio systems intentionally transform the source. None of these realities contradict the lossless claim. They simply exist outside it.

The listener who wishes to minimize variables chooses a platform that prioritizes genuine studio masters, disables loudness normalization, ensures that playback operates in exclusive or bit-perfect mode whenever possible, and routes the decoded PCM stream to a DAC without unnecessary intervention. At that point, the differences between platforms narrow considerably. What remains reflects primarily the quality of the source material supplied by the label and the engineering decisions made downstream of the file itself.

The bits are honest. The question is always which bits, processed how, and by whom, before they reach you.

The engineering challenge is not delivering the bits. Modern digital audio solved that problem decades ago. The challenge is preserving the integrity of those bits from the chosen master to the analogue output without unnecessary intervention along the way. That is ultimately the lesson of high-resolution streaming. The technology works. The question is whether everything surrounding it does.