Paper deep dive
How Much AI Is in This Track? Quantifying the Proportion of AI-Generated Stems in Hybrid Music Mixtures
Fernando Garcia de la Cruz, David LĂłpez-Ayala, Pablo Zinemanas, Emilio Molina, MartĂn Rocamora
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:AI-generated music is increasingly used at the stem level, with producers integrating synthetic drums, basslines, or vocals alongside human-performed instruments. However, current AI music detection systems are binary, treating tracks as either fully AI or fully human. In this paper, we reformulate AI music detection as a regression problem on a continuous AI energy ratio, alpha in [0, 1]. We propose a methodology that leverages a multi-track music dataset to assemble mixtures of human-performed and AI-reconstructed stems (obtained using a neural audio codec) with known proportions of each content type. Using this approach, we first show that a CNN-based model trained on fully AI-generated or human-performed tracks, which achieves >99% accuracy as a binary detector, when faced with mixed content, yields an output that rises with the AI stems' energy contribution, acting as a noisy and miscalibrated estimator. Our analysis of the influence of different stems shows that detection sensitivity depends on the instrument and reflects its frequency content: drums and guitar carry strong codec-artifact signatures, while vocals and bass are less detectable. Based on these insights, we train a similar CNN-based model for regression of alpha, achieving MAE = 0.076 and R^2 = 0.85 on held-out mixtures from the same pipeline. These results suggest that the regression formulation is an initial promising step towards AI-music detection in realistic music production workflows.
Tags
Links
- Source: https://arxiv.org/abs/2608.07285v1
- Canonical: https://arxiv.org/abs/2608.07285v1
Trouble viewing inline? Open PDF directly â
Full Text
37,116 characters extracted from source content.
Expand or collapse full text
HOW MUCH AI IS IN THIS TRACK? QUANTIFYING THE PROPORTION OF AI-GENERATED STEMS IN HYBRID MUSIC MIXTURES Fernando Garcia de la Cruz 1 David LĂłpez-Ayala 1 Pablo Zinemanas 2 Emilio Molina 2 MartĂn Rocamora 1 1 Music Technology Group, Universitat Pompeu Fabra, Barcelona, Spain 2 BMAT Licensing S.L., Barcelona, Spain fergarciadlc@gmail.com ABSTRACT AI-generated music is increasingly used at the stem level, with producers integrating synthetic drums, basslines, or vocals alongside human-performed instruments. However, current AI music detection systems are binary, treating tracks as either fully AI or fully human. In this paper, we reformulate AI music detection as a regression prob- lem on a continuous AI energy ratio, α â [0, 1]. We propose a methodology that leverages a multi-track music dataset to assemble mixtures of human-performed and AI- reconstructed stems (obtained using a neural audio codec) with known proportions of each content type. Using this approach, we first show that a CNN-based model trained on fully AI-generated or human-performed tracks, which achieves > 99% accuracy as a binary detector, when faced with mixed content, yields an output that rises with the AI stemsâ energy contribution, acting as a noisy and miscali- brated estimator. Our analysis of the influence of different stems shows that detection sensitivity depends on the in- strument and reflects its frequency content: drums and gui- tar carry strong codec-artifact signatures, while vocals and bass are less detectable. Based on these insights, we train a similar CNN-based model for regression of α, achieving MAE = 0.076 and R 2 = 0.85 on held-out mixtures from the same pipeline. These results suggest that the regression formulation is an initial promising step towards AI-music detection in realistic music production workflows. 1. INTRODUCTION The prevalence of AI-generated content is growing rapidly. Industry reports estimate that fully AI-generated tracks now account for more than 50% of daily uploads on ma- jor streaming platforms [1]. Industry studies also report that 97% of listeners cannot reliably distinguish fully AI- generated tracks from human-made recordings [2]. This raises concerns about transparency and fair remuneration for artists and highlights the need for automated detection © F. Garcia de la Cruz. Licensed under a Creative Com- mons Attribution 4.0 International License (C BY 4.0). Attribution: F. Garcia de la Cruz, âHow Much AI Is in This Track? Quantifying the Pro- portion of AI-Generated Stems in Hybrid Music Mixturesâ, in Proc. of the 27th Int. Society for Music Information Retrieval Conf., Abu Dhabi, UAE, 2026. tools that are also aligned with disclosure obligations, such as those in Article 50 of the EU AI Act [3]. Current AI music detection systems achieve near- perfect accuracy (>99%) when classifying complete tracks as fully AI-generated or human-produced [4, 5]. These detectors exploit architectural artifacts: the characteristic spectral peaks introduced by transposed convolution layers in neural audio codec decoders [6]. Yet, it is unclear how this binary paradigm applies when AI-generated stems are combined with human performances on a single track [7]. In this work, we reformulate AI music detection as a regression problem on a continuous AI energy ratio α â [0, 1], representing the proportion of a mixâs total acous- tic energy contributed by AI-generated stems. Studying this reformulation requires controlled mixtures of human- performed and AI-generated stems with known α values. To enable this, we propose a methodology for construct- ing hybrid mixtures from multi-track music datasets based on three components: (i) encodeâdecode reconstruction of human-performed stems through a neural audio codec, which imprints the same architectural artifacts that binary detectors learn to detect while preserving the original mu- sical content; (i) combinatorial stem replacement, which for a track with n stems produces all 2 n real/AI configura- tions, ensuring dense sampling of α; and (i) evaluation at the level of the AI energy ratio α, which correctly weights each stem by its acoustic contribution. We apply this methodology using the MoisesDB multi- track dataset [8] and the EnCodec neural audio codec [9], producing controlled mixtures across 240 professionally produced tracks. Our experiments reveal that a state-of- the-art binary detector [4], which achieves > 99% ac- curacy on pure tracks, produces an output score that in- creases with α, effectively acting as a noisy and miscali- brated estimator of the AI energy ratio. Detailed analysis of the contribution of the different stems to the output indi- cates that the detection sensitivity is frequency-dependent: while drums and guitars are more reliably identified, vo- cals are only partially detected, and bass remains largely undetected. Building on these findings, we train a CNN regressor using the same architecture to estimate α di- rectly, achieving a Mean Absolute Error (MAE) of 0.076 (R 2 = 0.85) on held-out mixtures from the same recon- struction pipeline. The AI stems in our study are codec reconstructions of arXiv:2608.07285v1 [eess.AS] 7 Aug 2026 real recordings, not generator output, so what we measure is the proportion of codec-generated energy under con- trolled conditions (Section 6). The binary detector is re- produced from Afchar et al. [4] as a diagnostic probe; our contributions are the reformulation, the methodology and released library, the characterization of binary-detector be- havior on hybrid mixtures, and a regression baseline. 2. RELATED WORK Neural audio codecs [9,10] enable high-fidelity audio com- pression into discrete token sequences, and audio and mu- sic generation systems [11, 12] leverage these representa- tions to generate music autoregressively from text or audio prompts. Other systems generate in the continuous latent space of an audio autoencoder. Stable Audio [13] synthe- sizes full tracks from text with latent diffusion, and Diff- A-Riff [14, 15] uses the same paradigm to generate one instrument stem at a time, conditioned on a reference mix through audio or text embeddings, while LiveBand [16] generates accompaniment causally with an adversarially trained transformer. The latter two produce hybrid tracks directly. Whatever the conditioning, all of these systems synthesize the waveform through a learned decoder built from transposed convolution layers. These layers intro- duce periodic spectral artifacts [17, 18], and current AI- music detectors learn to exploit them [6]. Early work on AI music detection focused on binary classification of complete tracks. Afchar et al. [4, 6] pro- posed a CNN-based detector that reaches 99.8% accu- racy on controlled datasets by exploiting codec-induced spectral artifacts. Rahman et al. [5] proposed SpecTT- Tra, a transformer with spectro-temporal tokenization that reaches an F1 score of 97% on full-track detection. How- ever, these detectors face important limitations. General- ization difficulties across generators were already noted by Afchar et al., and Cros Vila et al. [19] further demonstrated that binary detectors exhibit strong platform-specific bi- ases: a detector trained on music from one generation plat- form fails to generalize to another. What these detectors learn is therefore specific to a generator, not a universal sig- nature of AI-generated audio. Afchar et al. [4] also showed that simple audio transformations, such as pitch shifting or re-encoding, significantly degrade detection accuracy, ex- posing vulnerability to post-processing. Detection also de- grades on broadcast audio, where music is often short and mixed with speech; F1 scores fall below 60% [20]. A similar redefinition of the detection of AI-generated content has already taken place in adjacent fields. In vi- sual deepfake detection, recent work reformulates the task as multi-label classification over individual facial compo- nents or predicts pixel-level masks that localize manipu- lated regions within video frames [21, 22]. While existing detectors achieve high accuracy (>99%) on binary classification of fully AI-generated tracks, to our knowledge no prior work has examined detection when AI- generated content is mixed with human-created content at varying proportions. This setting is increasingly relevant as professional production integrates AI-generated com- ponents with human performances [7], a practice that the stem- and accompaniment-generation models discussed above directly support. Yet detectors trained under the as- sumption of either fully AI or fully human music can ex- hibit degraded performance on mixed content (Section 5). Motivated by this gap, and supported by the shift toward partial detection in adjacent fields, we study proportional AI content detection at the stem level and propose a re- gression baseline for the task. 3. PROBLEM FORMULATION 3.1 Binary AI-music Detection In prior work, AI music detection has been formulated as a binary classification problem [4]. Given an audio signal x â R TĂC (where T is the number of samples and C is the number of channels), the task is to predict y â 0, 1, where y = 0 indicates human-produced music and y = 1 indicates AI-generated music. The detectors identify artifacts in the form of spectral peaks introduced by the transposed convolution layers of the decoder, the audio counterpart of the checkerboard ar- tifacts of image generation [17]. Afchar et al. [6] call this signature an artifact fingerprint, and in their released code, a fakeprint, the term we use here. The peaks appear at f artifact = f s /(2Ă stride), where f s is the sampling rate. Because the strides are fixed by the architecture, the peaks appear at the same frequencies in all audio the decoder generates. 3.2 Mixed-Content Production Scenarios Assuming a linear mixture, a track is the sum of n stems, x mix = P n i=1 s i , where each stem s i is either human- produced (s i â H ) or AI-generated (s i â A). To quantify the AI content of a mix, we define the AI energy ratio α = P s i âA RMS(s i ) 2 P n i=1 RMS(s i ) 2 ,(1) where RMS(·) denotes the root-mean-square waveform amplitude values of a stem and α is the fraction of total signal energy contributed by AI-generated stems. Mixing alters artifact detectability. The spectral peaks are present only in the AI stems, so combining them with human content in overlapping frequency bands reduces their prominence relative to the surrounding spectrum. 3.3 Proportional Detection via Regression We reformulate AI music detection as a regression prob- lem. Given a mixed audio signal x mix , the task is to predict Ëα = f Ξ (x mix )â [0, 1], where Ëα is the estimated AI energy ratio and f Ξ is a learned function parameterized by Ξ. We report mean absolute error (MAE) and root mean squared error (RMSE) in the units of α, the coefficient of determination (R 2 ), and Pearsonâs r. A proportional output such as âapproximately X% AI-generatedâ carries more information than a binary label and better reflects hy- brid production workflows. 3.4 Key Challenges This reformulation introduces two challenges beyond bi- nary classification. Regression must distinguish arbitrary values of α, and the same α can arise from different stem combinations (50% AI vocals versus 50% AI drums), so detectability depends on how much energy the AI stems contribute and on where in the spectrum that energy lies. 4. METHODOLOGY We introduce a methodology for constructing controlled hybrid mixtures of human-performed and AI-generated stems with a known AI energy ratio α, as defined in Eqn (1). Note that the AI-generated stems in our study are codec reconstructions of real recordings. The methodology is guided by three principles: (1) stem-level granularity, reflecting the modular structure of modern production; (2) artifact isolation through recon- struction, which reduces content confounds between AI and human conditions; and (3) combinatorial complete- ness, which provides dense coverage of α over the full [0, 1] range. The methodology is agnostic to both the source corpus (any multi-track dataset with isolated stems can serve as input) and the codec used for reconstruction. 4.1 AI Reconstruction Pipeline We build AI stems by encoding and decoding existing human-performed stems through a neural audio codec. This encodeâdecode process introduces the characteristic spectral artifacts of the decoderâs transposed convolution layers [6] while preserving the original musical content. This isolates the detection problem to architectural arti- facts, eliminating confounds from differences in musical style or content between the AI and human conditions. The process is deterministic and therefore reproducible. The methodology is also codec-agnostic; any neural codec ex- posing encode and decode operations can be substituted. We instantiate the pipeline with EnCodec [9], recon- structing stems at a single bitrate of 3 kbps. When repro- ducing the evaluation conditions in [4], the reconstructed stems are stored as MP3 at 256 kbps, 44.1 kHz, so that any observed detection failure on mixed content is attributable to the mixing process and not to a format mismatch. 4.2 Combinatorial Mix Generation Given a track with n stems indexed by I =0, 1,...,nâ 1, we define a mix specification by the subset S â I of stems replaced with their AI-reconstructed versions. The resulting mix is x mix (S) = X iâI s real i + X iâS s AI i ,(2) where s real i and s AI i denote the authentic and AI- reconstructed versions of stem i. Stems are summed with equal weights, preserving the original energy relationships defined by the source mix. We organize mixes into sets indexed by the number of AI-reconstructed stems: U k =x mix (S) : S â I, |S| = k, k = 0, 1,...,n. (3) Each set U k contains n k mixes, for a total of 2 n mixes per track. The ratio k/n, the stem fraction, counts how many stems are replaced, whereas the target variable α (Eqn (1)) weights each stem by its energy. Sets U 0 (all real) and U n (all AI) correspond to the bi- nary detection setting; the intermediate sets U 1 ,...,U nâ1 contain the mixed-content cases. Within a set, mixes share the stem fraction k/n but span a range of α values and dif- fer in which instruments are AI. This last property is what enables the instrument-level analysis of Section 5.3. 4.3 Multi-track audio dataset The methodology applies to any multi-track dataset pro- viding isolated stems, including those developed for source separation (e.g., MUSDB18 [23], Slakh [24]). We instan- tiate the methodology on MoisesDB [8], a professionally mixed multi-track dataset providing high-quality stem iso- lations across diverse genres. We use 240 tracks, with stem counts ranging from 2 to 10 (most tracks contain 4â 7 stems) categorized by instrument type (vocals, drums, bass, guitar, and other instruments such as keyboards, strings, and synthesizers). The variable stem count is a fea- ture of the design: tracks with different n produce different sets of achievable α values, and aggregated across the cor- pus, these provide dense coverage of the full [0, 1] range, with most mixes concentrated around mid-range values. In total, this yields 21,212 unique mixes across the 240 tracks. We release a reproducible Python library 1 that con- structs mixtures from a local MoisesDB installation. 5. EXPERIMENTS AND RESULTS We apply the methodology of Section 4 to construct hy- brid mixtures and compare two models on them: the re- produced binary detector, and a regression model with the same backbone trained on α. 5.1 Experimental Setup Models. Following [4], we implement a CNN classifier: a six-block convolutional backbone operating on magni- tude spectrograms (STFT with 2048-sample window, 512- sample hop, frequencies above 16 kHz discarded), fol- lowed by global average pooling and a dense head. We train two variants of this architecture that differ only in their output layer and loss. The binary detector uses a sigmoid-activated classification head trained with binary cross-entropy on the original 0, 1 labels. The regres- sion model replaces the classification head with a single- unit, sigmoid-activated regression head trained with mean squared error against the AI energy ratio α (Eqn (1)). At inference, both models produce per-window scores Ëy â 1 ARIA (Authentic vs. Reconstructed Isolated Audio); code available at https://github.com/fergarciadlc/aria. [0, 1], which are averaged across non-overlapping windows to obtain a track-level score. Training.The binary detector is trained on FMA- Medium [25] and its EnCodec-reconstructed counterpart at 3 kbps, following the original methodology by Afchar et al. [4]. The regression model is trained on the hybrid mixtures produced by our methodology applied to Moi- sesDB at the 3 kbps EnCodec condition (Section 4.3), with α computed per 5-second window as the regression target. We initially trained both models on 1-second windows, matching prior work [4]. The regression model performed better with 5-second windows, particularly near the bound- aries of the α range where short windows produced nois- ier estimates, so we retrained the binary detector at 5- second windows as well and report the matched compar- ison throughout. Section 5.5 reports both window lengths. Train, validation, and test splits are performed at the track level, stratified by stem count, so that all mixes and all windows from a given track land in the same partition, pre- venting content and temporal-correlation leakage. The 240 tracks are split 80/10/10, leaving 24 held-out tracks and 1,792 test mixtures. Reproduction validation.The binary detector achieves 99.97% accuracy and 99.98% F1 on a held-out FMA evaluation set, closely matching [4]. On the pure endpoints U 0 and U n of our mixture corpus, the detector assigns scores near 0 and 1, respectively, confirming that the reconstruction pipeline preserves the codec artifacts on which the binary classifier relies. Evaluation. We run inference on the held-out test split and report track-level scores, obtained by averaging the per-window scores of each mixture; the instrument-level analysis of Section 5.3 uses window scores directly. 5.2 Energy-Driven Detection Response We first examine how each modelâs output Ëy depends on the AI energy ratio α (Figure 1, medians binned in 10% increments, both models at 5-second windows). The binary detectorâs output increases with α, but the increase is not linear: its median stays near 0.05 for α < 0.5 and then rises to⌠0.97 for α > 0.9. The detector was trained only on the endpoints α â 0, 1 and has never seen mixed content, so this response shows that codec ar- tifacts in hybrid mixtures carry proportional information about AI content, which the binary classifier picks up inci- dentally as a noisy and miscalibrated estimator of α. The regression model, trained directly on α, follows the diagonal across the full range, while the binary detec- tor underestimates AI content in the low and mid range: at α â [0.4, 0.5] its median is 0.10, against 0.44 for the regression model. Part of this gap comes from window length (Section 5.5). The proportional signal is therefore present in mixed content, and the training objective deter- mines how well it is recovered. 5.3 Instrument-Level Detection Sensitivity A natural place to look for the source of detection sensitiv- ity is the stem level, the unit at which AI-generated content 0.0-0.10.1-0.20.2-0.30.3-0.40.4-0.50.5-0.60.6-0.70.7-0.80.8-0.90.9-1.0 AI energy ratio α 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Median track-level score Ì y Perfect prediction Binary CNN (5s) Regression CNN (5s) Figure 1. Median Ëy as a function of the AI energy ratio α for the binary and regression models, both at 5-second windows; error bars show the interquartile range. is integrated into modern productions. We isolate each in- strumentâs contribution to detection. We select tracks in our test corpus whose stems are exactly vocals, drums, bass, guitar, the four most common stem types in the cor- pus (instantiated from [8]). Each of these appears individu- ally in over 90% of tracks, whereas the next most common type, piano, is present in only 46%. For each selected track we generate all 16 possible AI/real configurations, exclud- ing the all-real case. For each instrument, we compare Ëy when that stem is real versus AI-generated, marginalizing over the other three stems. Figure 2 shows the resulting score distributions by window for both models, at 1-second windows. This is the setting in which the binary detector tracks α most closely (Table 1), and therefore the most fa- vorable one for it. For the binary detector (top row), drums and guitar sep- arate most clearly, with the AI distribution concentrated near 1 and the real distribution near 0; whereas the distri- butions for vocals and bass substantially overlap. Sensi- tivity therefore depends on the instrument, and Section 5.4 traces this to where each stemâs energy lies in the spectrum. Notably, vocals and bass become partially separable under regression (bottom row), indicating that the regressor can extract proportional information from a weaker fakeprint signal even when the binary detector cannot. 5.4 Fakeprint Analysis To verify that the per-instrument differences in detection observed in Section 5.3 reflect the underlying artifact dis- tribution rather than other detectorâs factors, we compute the fakeprints directly from the stems. Following Afchar et al. [6], we obtain a fakeprint by time-averaging the log- power spectrum of a stem and subtracting the local minima over sliding frequency windows, keeping only the positive residual peaks. Theory predicts that AI-generated content exhibits peaks at frequencies determined by the strides of the codecâs transposed convolution layers [6]; whereas real recordings do not present these artifacts. We follow the original procedure but restrict the bandwidth to [0, 5] kHz to capture the lower half of the audible range where instru- ment fundamentals and lower harmonics live; at this band- 0 2 4 Binary Density GuitarDrumsVocalsBass Real AI 0.00.51.0 0 1 Regression Density 0.00.51.00.00.51.00.00.51.0 Window score Figure 2. Kernel density estimates of Ëy for each instru- ment under real and AI conditions, for the binary detector (top row) and regression model (bottom row), both at 1- second windows. 0.0 0.1 0.2 Norm. peak residue BassDrums Real AI 010002000300040005000 Frequency (Hz) 0.0 0.1 0.2 Norm. peak residue Guitar 010002000300040005000 Frequency (Hz) Vocals Figure 3. Average fakeprint curves (Gaussian-smoothed) for Real and AI stems across bass, drums, guitar, and vo- cals over the [0, 5] kHz band, highlighting stem-specific spectral differences between real and generated content. width and after Gaussian smoothing, the curves capture broad spectral envelope differences between real and AI conditions rather than the narrow stride-determined peaks predicted by theory. Following the same marginalization strategy as Section 5.3, we pool fakeprints from all mixes in which a given stem is AI-generated versus real and average within each group. Figure 3 shows the average fakeprints for the four different types of stems. The fakeprint differences exhibit the following patterns. Bass shows essentially no separation between real and AI conditions, whereas vocals show very subtle differences. In contrast, drums show clear separation from⌠2500 Hz onward, with the AI curve lifted above the real curve. Gui- tar shows a distinctive crossover: the AI curve sits below the real curve in the low-mid range (⌠500â1500 Hz) and substantially above it from⌠2500 Hz onward, suggesting that AI-reconstructed guitar both attenuates low-mid spec- tral structure and introduces high-frequency artifact peaks. These results align with the binary detectorâs instrument sensitivity reported in Section 5.3, indicating that the vari- ation is rooted in the codecâs artifact distribution: bass and vocal evade detection because the codec leaves a lower de- tectable artifact signature in their content, while drums and ModelMAE RMSER 2 Pearson Binary CNN, 1 s (α-est.)0.2280.289 â0.1930.617 Binary CNN, 5 s (α-est.)0.2890.360 â0.8510.505 Regression CNN, 1 s0.1040.1240.7820.892 Regression CNN, 5 s0.076 0.102 0.8540.925 Table 1. Proportional-detection performance on the held- out test split, computed on per-mixture scores, each aver- aged across the windows of the mixture, at 1- and 5-second windows. The binary detectorâs raw fake probability is used as an implicit α estimate. guitar carry strong signatures that the detector can identify. 5.5 Regression Baseline Section 5.1 describes the training setup; here we report per- formance on the held-out test split. Baseline. As an implicit baseline we use the binary de- tector directly as an α estimator, treating its fake probabil- ity Ëy as a prediction of α. This is an unfair comparison, since the binary detector was not trained for this objective, but it quantifies how much information about α a binary classifier captures incidentally, which is exactly the ques- tion raised by Section 5.2. Results. Table 1 reports performance on the test split. At the matched 5-second window, the regression model achieves MAE = 0.076, RMSE = 0.102, R 2 = 0.85, and Pearson r = 0.93. The binary detectorâs implicit estimate reaches MAE = 0.289 and R 2 = â0.85. A negative R 2 means the estimate is less accurate than always predict- ing the dataset mean, even though the same detector clas- sifies the pure endpoints with > 99% accuracy. Training directly on α therefore gives a large gain over reusing a bi- nary score. Figure 4 shows predicted vs. true α averaged across the windows at the track level. Window length. Table 1 also reports both models at 1-second windows, and the two objectives respond to win- dow length in opposite directions. The regression model improves with longer windows (MAE 0.076 vs. 0.104, R 2 0.85 vs. 0.78), since a continuous quantity is estimated more stably over more content; this motivates the 5-second choice. The binary detector, which is never trained on α, gets worse (R 2 â0.19 vs. â0.85), since averaging over a longer window blurs the weak proportional signal it cap- tures. The regression model is more accurate at both win- dow lengths, and the matched 5-second setting is the one we compare on. Boundary compression. The regression model ex- hibits mild compression near the boundaries: predictions floor at ⌠0.05â0.10 for fully real mixtures and ceil- ing at ⌠0.85â0.95 for fully AI mixtures, explaining the convergence with the binary detector at α = 1 noted in Section 5.2. This is consistent with bounded sigmoid regression under MSE loss.Absolute errors neverthe- less remain low at the boundaries; the largest errors oc- cur in the mid range, where mixtures admit the most var- ied stem compositions and the analyses of Section 5.3 and Section 5.4 predict the greatest variance in detectability. 0.00.10.20.30.40.50.60.70.80.91.0 True AI energy ratio α 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Predicted α Perfect prediction Figure 4.Predicted versus true AI energy ratio α for the regression model, averaged across the windows of each mixture, on the held-out test split. 6. LIMITATIONS Our empirical findings are based on a single neural au- dio codec (EnCodec [9]) at a single bitrate (3 kbps). The AI stems are codec reconstructions which isolate codec artifacts from content confounds and enable controlled, reproducible experiments, but bound the scope of our claims. Whether these findings transfer to other codecs (e.g., DAC [10]) or to commercial generators such as Suno and Udio, whose artifacts may lie at different frequencies, remains untested. That said, the artifacts are a property of the decoder architecture [6], and our methodology does not depend on which architecture produces them. Our mixtures are also constructed without the post- processing typical of professional production workflows (EQ, compression, level balancing, creative editing of AI stems), which could alter both detection accuracy and the effectiveness of the AI energy ratio estimation. Accord- ingly, we view this work as a controlled methodological foundation rather than a deployment-ready detector. Trans- lating these findings into production tools requires corpora in which human-performed stems are paired with stems from actual music generators, for which stem-level access is currently limited. Two routes could supply them: collab- oration with music platforms holding stem-level catalogs, and stem-conditioned generative models [14, 16], which generate stems that fit existing human-performed material. The methodology and library we release are codec- and corpus-agnostic and can be instantiated on such data as it becomes available. 7. CONCLUSION In this work, we address AI-music detection on mixed con- tent in a controlled setting by reformulating the task as re- gression on a continuous AI energy ratio α and proposing a codec- and corpus-agnostic methodology for construct- ing hybrid mixtures with known α from any multi-track corpus. Applying this methodology to MoisesDB and En- Codec, we show that codec artifacts in hybrid mixtures carry proportional information about α: a binary detector trained only on α â 0, 1 captures this signal inciden- tally as a noisy estimator, whereas a regression model with the same backbone trained directly on α extracts it substan- tially more accurately (MAE = 0.076, R 2 = 0.85). Detec- tion sensitivity depends on the stem: some carry stronger codec-artifact signatures than others, in line with their fre- quency content, as the fakeprint analysis confirms. The proportional-detection formulation offers a more informative target than binary classification for the real- istic case of hybrid music production, and the methodol- ogy can be instantiated with other codecs and with stems from commercial AI generators. Because sensitivity de- pends on the stem, instrument-aware detection is a natural next step: applying detection to separated stems before ag- gregation, or conditioning on stem identity during training. Both models also produce per-window scores, so reporting them over time gives a route to temporal localization; de- tecting generated material within a stem, such as a single synthetic drum hit, would require mixtures built at finer granularity than the stem. The methodology and accom- panying library are released as a foundation for continued research on hybrid AI-music detection. 8. ACKNOWLEDGMENTS This work is supported by the âCĂĄtedra IA y MĂșsicaâ project (TSI-100929-2023-1),funded by the Sec- retarĂa de Estado de DigitalizaciĂłn e Inteligencia Artificial, the European Union-Next Generation EU funds and BMAT Music Innovators.And by the âIMPAâ project (PID2023-152250OB-I00) funded by MCIU/AEI/10.13039/501100011033/FEDER, UE. 9. REFERENCES [1] J. Wendel, âDeezer: AI music has surpassed 50% of new music uploads for the first time,â Deezer News- room, Jul. 2026, accessed: July 29, 2026. [Online]. Available:https://newsroom-deezer.com/2026/07/ ai-music-exceeds-50-percent-daily-uploads-deezer/ [2] â, âDeezer/Ipsos survey: 97% of people canât tell the difference between fully AI-generated and human made music â clear desire for trans- parency and fairness for artists,â Deezer News- room, Nov. 2025, accessed: July 29, 2026. [On- line]. Available: https://newsroom-deezer.com/2025/ 11/deezer-ipsos-survey-ai-music/ [3] X. Serra, R. O. Araz, R. Batlle-Roca, L. Juvela, D. LĂłpez, and M. Rocamora, âTechnical solutions for marking and detecting AI-generated audio content in the context of article 50(2) AI act,â EuropeanCommission,Directorate-Generalfor Communications Networks, Content and Technology, Final Study Report, 2026. [Online]. Available: https: //op.europa.eu/en/publication-detail/-/publication/ 4f7b8585-4829-11f1-8095-01a75ed71a1/ language-en [4] D. Afchar, G. Meseguer-Brocal, and R. Hennequin, âAI-generated music detection and its challenges,â in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, p. 1â5. [5] M. A. Rahman, Z. I. A. Hakim, N. H. Sarker, B. Paul, and S. A. Fattah, âSONICS: Synthetic or not â identifying counterfeit songs,â in Proceedings of the International Conference on Learning Representations (ICLR), 2025. [Online]. Available: https://openreview. net/forum?id=PY7KSh29Z8 [6] D. Afchar, G. Meseguer-Brocal, K. Akesbi, and R. Hennequin, âA Fourier explanation of AI-music ar- tifacts,â in Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2025, p. 739â746. [7] J. Pons, Z. Zukowski, J. D. Parker, C. J. Carr, J. Taylor, and Z. Evans, âMusic and artificial intelligence: Artis- tic trends,â 2025, arXiv:2508.11694. [8] I. Pereira, F. AraĂșjo, F. Korzeniowski, and R. Vogl, âMoisesDB: A dataset for source separation beyond 4-stems,â in Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2023, p. 619â626. [9] A.DĂ©fossez,J.Copet,G.Synnaeve,and Y. Adi,âHigh fidelity neural audio compres- sion,âTransactions on Machine Learning Re- search (TMLR), 2023. [Online]. Available:https: //openreview.net/forum?id=ivCd8z8zR2 [10] R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, âHigh-fidelity audio compression with im- proved RVQGAN,â in Advances in Neural Informa- tion Processing Systems (NeurIPS), 2023, p. 27 980â 27 993. [11] Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Sharifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasacchi, and N. Zeghidour, âAu- dioLM: A language modeling approach to audio gen- eration,â IEEE/ACM Transactions on Audio, Speech, and Language Processing (TASLP), vol. 31, p. 2523â 2533, 2023. [12] J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Syn- naeve, Y. Adi, and A. DĂ©fossez, âSimple and control- lable music generation,â in Advances in Neural In- formation Processing Systems (NeurIPS), 2023, p. 47 704â47 720. [13] Z. Evans, C. J. Carr, J. Taylor, S. H. Haw- ley, and J. Pons, âFast timing-conditioned latent audio diffusion,â in Proceedings of the Interna- tional Conference on Machine Learning (ICML), 2024,p.12 652â12 665.[Online].Available: https://proceedings.mlr.press/v235/evans24a.html [14] J. Nistal, M. Pasini, C. Aouameur, M. Grachten, and S. Lattner, âDiff-A-Riff: Musical accompaniment co- creation via latent diffusion models,â in Proceedings of the International Society for Music Information Re- trieval Conference (ISMIR), 2024, p. 272â280. [15] J. Nistal, M. Pasini, and S. Lattner, âImproving mu- sical accompaniment co-creation via diffusion trans- formers,â in Proceedings of the Audio Imagination Workshop at the Conference on Neural Information Processing Systems (NeurIPS), 2024. [Online]. Avail- able: https://openreview.net/forum?id=zyE3Kdd85t [16] M. Pasini, J. Nistal, M. R. Bjare, S. Lattner, and G. Fazekas, âLiveBand: Live accompaniment gener- ation in the audio domain,â 2026, arXiv:2606.03803. [17] A. Odena, V. Dumoulin, and C. Olah, âDeconvolution and checkerboard artifacts,â Distill, vol. 1, no. 10, 2016. [Online]. Available:https://distill.pub/2016/ deconv-checkerboard/ [18] J. Pons, S. Pascual, G. Cengarle, and J. SerrĂ , âUpsam- pling artifacts in neural audio synthesis,â in Proceed- ings of the IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2021, p. 3005â3009. [19] L. Cros Vila, B. L. T. Sturm, L. Casini, and D. Dal- mazzo, âThe AI music arms race: On the detec- tion of AI-generated music,â Transactions of the In- ternational Society for Music Information Retrieval (TISMIR), vol. 8, no. 1, p. 179â194, 2025. [20] D. LĂłpez-Ayala, A. Cabello, P. Zinemanas, E. Molina, and M. Rocamora, âAI-generated music detection in broadcast monitoring,â in Proceedings of the IEEE In- ternational Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2026, p. 12 342â12 346. [21] L. Deng, Y. Zhu, D. Zhao, and F. Chen, âA multi-label classification method based on transformer for deep- fake detection,â Image and Vision Computing, vol. 152, p. 105319, 2024. [22] J. Xu, X. Liu, W. Lin, W. Shang, and Y. Wang, âLocal- ization and detection of deepfake videos based on self- blending method,â Scientific Reports, vol. 15, no. 1, p. 3927, 2025. [23] Z. Rafii, A. Liutkus, F.-R. Stöter, S. I. Mimilakis, and R. Bittner, âThe MUSDB18 corpus for music separa- tion,â Zenodo, Dec. 2017. [24] E. Manilow, G. Wichern, P. Seetharaman, and J. Le Roux, âCutting music source separation some Slakh: A dataset to study the impact of training data quality and quantity,â in Proceedings of the IEEE Workshop on Applications of Signal Processing to Au- dio and Acoustics (WASPAA), 2019, p. 45â49. [25] M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, âFMA: A dataset for music analysis,â in Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2017, p. 316â323.