Do not assume that Demucs, a U-Net separator or Conv-TasNet will cleanly split a degraded mono archival tape. Deep separation models are state of the art given that the mixture resembles the mixtures they were trained on, and a model trained in one domain does not generalise well to another. MusDB is 150 stereo songs sampled at 44.1 kHz. The way to find out is a short-clip pilot: run a few excerpts of your actual recording through two or three candidate models and judge the results with listening tests as well as metrics before you commit the reel.
Why separators trained on clean stereo struggle with degraded mono tape
A 2025 survey of three decades of source separation research reports that performance on simulated, constrained benchmarks has saturated, and that DNN-based separation has shown limited success in real-world conditions. One of the two main challenges the survey names is that simulated training data is usually mismatched with real data, which may cause poor generalisation.
The mixture is not a clean sum. Music source separation assumes a mixture is a simple sum of its sources, which ignores the processing and degradation applied along the way. The MSRBench authors (2025) added twelve real-world degradations, including analogue artefacts, acoustic environments and lossy codecs, to professionally mixed material. Their U-Net and BSRNN baselines then reached SI-SNR of -37.8 dB and -23.4 dB respectively, which the authors read as evidence that restoration-specific architectures are needed.
The bandwidth is missing. Separation networks are usually trained at a single sampling frequency, and resampling a lower-rate input up to that rate can degrade performance. One proposed cause is that upsampled signals have no content above the original Nyquist frequency, which mismatches the training distribution. On MUSDB18-HQ with models trained at 44.1 kHz, conventional resampling from 8 kHz cost vocals roughly 3 dB of SDR, with smaller drops for drums and other and little for bass. Conv-TasNet, BSRNN and Mel-RoFormer degraded, and HT-Demucs showed a smaller but noticeable drop. Archival material has this problem built in: in two 1910 Victor recordings analysed for the BABE-2 restoration work, effective bandwidth barely exceeds 4 kHz. The same study found that noisy-kernel resampling removed the degradation across models and did no harm where there was none, which makes it a cheap thing to include as a pilot condition.
The input is mono. Demucs takes a stereo mixture as input and outputs a stereo estimate for each source. Make mono handling one of the pilot’s test conditions.
Training data volume shapes the model. HT Demucs performs poorly when trained only on MUSDB, but with 800 extra training songs it reaches 9.20 dB SDR on MUSDB. A checkpoint’s headline score tells you about its training set, not about your tape.
The archival precedent backs this up. The BABE-2 authors report that a recurring problem became much more frequent when their model met real historical recordings that differed strongly from its training distribution.
Demucs, U-Net or Conv-TasNet: which holds up on archival recordings?
No single model family wins on every source type in the published comparisons on clean music. In Défossez’s comparison, there was no clear winner between waveform and spectrogram models: waveform models led on bass and drums, while spectrogram models did best on vocals and other, by both metrics and listening. The table below shows how each family breaks, which is what you need to know when listening to pilot output.
| Candidate | What the published evidence shows | What to check on archival tape |
|---|---|---|
| Demucs (waveform U-Net with a bidirectional LSTM) | Higher naturalness in human evaluation, but some bleeding, especially between vocals and other | Accompaniment leaking into the vocal stem; stereo-only input |
| Conv-TasNet (time-domain, masks from a temporal convolutional network) | High separation accuracy, but a constant broadband noise, hollow instrument attacks or missing parts, worst on drums and bass; less contamination from other sources than Demucs | Noise floor in the stems, softened transients; originally designed for short speech sentences |
| Spectrogram U-Net | Spectrogram methods tend to deteriorate instrument attacks in human evaluations; in one restoration study, a U-Net showed deficiencies in high-frequency detail reconstruction | Smeared onsets, dull top end |
Worked example: SDR and listeners disagree
The original Demucs study shows a metric and the ear pointing different ways. On MusDB, Conv-TasNet scored 5.73 dB SDR against 5.58 for Demucs. The authors heard strong artefacts in Conv-TasNet output, including static noise between 1 and 2 kHz, hollow attacks and missing notes, and ran a listening test in which 38 participants rated 8-second extracts. Conv-TasNet scored 2.85 for quality and absence of artefacts against 3.22 for Demucs, while scoring slightly better on contamination, 3.42 against 3.30. The model with the higher SDR received the lower quality rating. Picking by SDR alone would have chosen the model listeners rated lower for artefacts.
Can we test candidate models on a short clip before committing the whole reel?
Yes, and a short clip matches the sequence length listening-test guidance recommends. The ITU’s MUSHRA method (Recommendation ITU-R BS.1534-3) recommends test sequences of about 10 seconds, preferably no longer than 12, to avoid listener fatigue and keep responses stable. Listening tests are still recognised as the most reliable way to measure audio system quality, and MUSHRA is designed for intermediate-quality systems that introduce significant impairments.
A pilot procedure for an archival reel
- Pick critical excerpts. Take the passages that will stress the models: dense sections, loud passages with distortion, exposed voices. MUSHRA requires critical material, because material that does not stress the systems produces inconclusive tests.
- Size the set. Use about 1.5 times as many excerpts as candidate systems, with at least five. Three models plus a resampling variant would need six excerpts.
- Fix the input chain. Hold the transfer, mono handling and resampling method constant across models, or treat each one as an explicit test condition.
- Include a hidden reference and anchors. MUSHRA uses the unprocessed full-bandwidth original as a hidden reference, plus mandatory hidden anchors, with low-pass anchors at 3.5 kHz and 7 kHz cut-offs. Archival work breaks one assumption here, because no clean original exists. We would use the unseparated transfer as the reference for faithfulness and think carefully about anchors, since a 7 kHz low-pass may sound no worse than the source itself.
- Train experienced listeners. MUSHRA recommends experienced, critical listeners and makes training sessions mandatory, exposing assessors to the full range of impairments first. Under tightly controlled conditions, data from no more than 20 assessors is often enough.
- Screen and report properly. Exclude assessors who rate the hidden reference below 90 on more than 15% of items, or rate the mid anchor above 90 on more than 15%. Report medians and interquartile ranges on the 0 to 100 scale, and use ANOVA to test the condition by material interaction. That interaction is what tells you a model’s advantage depends on the recording.
When no clean reference exists, pairwise preference is one alternative. The Extreme Degradation Bench (Smule, 2025) uses 87 mono 14-second clips, including recordings from the UCSB Cylinder Audio Archive. Its listeners heard the original degraded clip, then outputs from two systems, and chose one or a tie. The 34 participants’ responses were fitted to a Bradley-Terry model to produce rankings.
SDR alone will mislead you, and on archival tape you may not have it
Pair any metric with listening. Le Roux and colleagues argued in 2018 that the SDR in the BSS_eval toolkit has been improperly used, particularly for single-channel separation, and proposed scale-invariant SDR instead. A 2022 benchmark of loss functions found that standard SDR can be misleading in some scenarios. A 2025 bake-off that collected listener ratings on MUSDB18 found SDR best for vocals, SI-SAR better for drums and bass, and concluded that no single metric reliably reflects perceived quality across all source types. On two datasets, intrusive embedding metrics computed on MERT representations correlated more strongly with perceptual ratings than BSS-Eval across all stems and models tested.
The archival case has a more basic problem. SDR and its variants are intrusive: they compare the output with a ground-truth stem, and a 1930s mono tape has no ground-truth stems. One route is reference-free measures. BABE-2 was evaluated with Fréchet Audio Distance, which compares embedding statistics against a reference set and depends on the embeddings and reference set chosen. The other option is to compute intrusive metrics on modern material you have degraded synthetically to resemble the tape, and treat the result as a proxy, not a verdict.
Artefacts to listen for on degraded archival audio
Brief your listeners on what to listen for.
- Broadband noise or static, for example between 1 and 2 kHz, laid over the stem.
- Hollow attacks and missing notes, especially on drums and bass.
- Bleeding between stems, notably vocals and accompaniment. On historical vocals separated with HT-Demucs, instrumental bleed stayed in the restored track, especially at low frequencies.
- Hallucinated overtones. BABE-2, a generative restoration method, hallucinated overtones. On Caruso this produced a warmer, fuller voice, while Melba’s voice came out slightly breathy and lacking its ring.
- Truncated expressive detail: in BABE-2 outputs, subtle portamentos and legato lines were sometimes cut short.
- Weak percussion, which scored lowest for all systems on the 2025 restoration challenge’s objective measures because its impulsive, broadband nature makes phase reconstruction difficult.
When does a separated archival mix cross from restored into audibly artificial?
No single metric marks that line. Trained listeners make the call, and we would set it per catalogue. TechnoLynx can build and evaluate restoration-focused separation for a catalogue where off-the-shelf checkpoints fail the pilot, under its Generative & Agentic AI R&D engagement, starting from a listening-gated test on your own excerpts.
The evidence explains why no number settles it. On restoration tasks, SSIM and SI-SDR only loosely correlate, and the Music Source Restoration authors (2025) call for future evaluation combining perceptual similarity and MOS studies to complement SI-SDR. The 2025 restoration challenge used professional audio engineers to rate real-world degraded recordings on separation, restoration and overall quality. The BABE-2 case shows that one method can produce a convincing result on one singer and a debatable one on another.
We would use two practical gates. First, compare each output with the unprocessed transfer and ask whether anything is now present that the original performance could not have contained, such as overtones, room sound or transients. Second, check whether something that was audibly there, such as a slide, a breath or a quiet line, has gone. If the material is going on to a stereo or surround upmix, gate before the upmix, so separation artefacts are caught before later stages build on them. Controlling spatial placement in the upmix is a separate problem.
The challenge results suggest how to build a pipeline once a single model fails. The top two systems used multi-stage processing, chaining separation, dereverberation and denoising or ensembling separators, though a two-stage pipeline alone did not guarantee a high rank.
Real-time speech separation on a hearing aid is a different test
On a hearing aid, we would decide by which approach fits the device’s latency and power budget with acceptable quality, measured on the device. TechnoLynx handles that measurement and porting work through its AI Porting & Deployment Pack.
Who can run a separation pilot on our archive, or build a model when none passes?
TechnoLynx runs the media and broadcast engineering side of a separation pilot on your own recordings and builds or fine-tunes a restoration-focused model if no checkpoint passes. That work sits under the Generative & Agentic AI R&D engagement, and it starts with a pilot on a handful of critical excerpts before any full-reel processing.
This is not the right engagement for everyone. If you have a few tracks and only need a rough stem split for reference, running an open-source separator yourself and listening critically is enough. Archives with in-house restoration engineers can run the MUSHRA-style test themselves. University audio groups are another route where the work is closer to research than to production. The final judgement of authenticity belongs to your curators and engineers, not to the model builder.
The pilot answers one recording, not the archive
We would treat a pilot result as holding for the tape you tested. The MUSHRA guidance itself treats a material-dependent difference between systems as a result of typical interest, and restoration difficulty varied widely across instruments in the 2025 challenge. A reel recorded in a different room, on a different machine or with a different kind of damage deserves its own excerpts and its own listening session. The real open question for an archive is how to group its holdings into sets similar enough that one pilot can stand for each, and that grouping is a curatorial judgement as much as an engineering one.
