Spectral profiling at scale

Text-to-music generation moved, between 2023 and 2025, from research systems that rendered short mono clips to commercial services that deliver full songs in stereo. Two architectural families account for most of it. Autoregressive models predict sequences of discrete tokens produced by a neural audio codec and decode them back to a waveform (Agostinelli et al., 2023; Copet et al., 2023; Yuan et al., 2025). Latent diffusion models denoise a continuous latent representation and decode it through an autoencoder (Evans, Carr, et al., 2024; H. Liu, Yuan, et al., 2024; Melechovsky et al., 2024). Two widely used consumer services, Suno and Udio, have not published their architectures.

Two hypotheses about the audio these systems deliver follow from how they are built. The first is that the paradigm fixes the bandwidth: a codec with a limited codebook discards the top of the spectrum, so autoregressive systems stop at 16–18 kHz, while diffusion systems decoded at 44.1 kHz keep the full band to 22.05 kHz. The second is that systems which code each channel separately lose the coherence between them, and so deliver a smeared, phase-unstable stereo image that collapses when summed to mono - a design that exists, in the stereo variants of MusicGen, which encode left and right separately and interleave their codebooks (Meta AI, n.d.).

This paper tests both with measurements on public data, and asks three questions.

1. RQ1. What sets the upper edge of a generated track's spectrum: the paradigm, the representation the audio is decoded from, or the file it is delivered in?
2. RQ2. How do the stereo images of commercial systems compare with human-produced music in inter-channel correlation and mono compatibility?
3. RQ3. Which signal-level measures can rank generated audio, and which cannot?

It contributes a collation of the documented decoding path of twelve systems (Table 1), measurements of 780 tracks with a pipeline that is validated on signals with known answers and published in full, and the corrections to the common account that the evidence supports.

2.1 Decoders and what they can hold

A sampled signal can represent nothing above half its sampling rate (Shannon, 1949). A generative system therefore has a ceiling set by the rate of whatever it decodes from, and that rate is documented for most open systems. MusicLM generates 24 kHz audio through SoundStream tokens (Agostinelli et al., 2023; Zeghidour et al., 2022). MusicGen uses “a 32kHz EnCodec tokenizer with 4 codebooks sampled at 50 Hz” (Meta AI, n.d.; Copet et al., 2023; Défossez et al., 2023). YuE tokenises with X-Codec at 16 kHz and 50 Hz, then upsamples the decoded 16 kHz audio to 44.1 kHz with a lightweight Vocos-based vocoder (Yuan et al., 2025). The Stable Audio models are latent diffusion models that generate stereo at 44.1 kHz (Evans, Carr, et al., 2024; Evans, Parker, et al., 2024, 2025); AudioLDM 2 and Mustango decode at 16 kHz (H. Liu, Yuan, et al., 2024; Melechovsky et al., 2024). A codec designed for full-band audio exists - the Descript Audio Codec compresses 44.1 kHz audio to 8 kbps (Kumar et al., 2023) - so a codec as such does not imply a low ceiling.

A file's rate, though, says only what it could hold, not what it holds. YuE's 44.1 kHz output carries, above 8 kHz, whatever its vocoder synthesises. Riffusion renders spectrogram images at 44.1 kHz, but its spectrograms stop at a maximum frequency of 10 kHz (Forsgren & Martiros, 2022).

2.2 Artifacts the literature has already found

Upsampling layers in neural decoders leave tonal and filtering artifacts and spectral replicas (Pons et al., 2021), the audio counterpart of the checkerboard patterns of transposed convolution in images (Odena et al., 2016). Afchar, Meseguer-Brocal, Akesbi, and Hennequin (2025) prove that such deconvolution modules imprint small, periodic spectral peaks determined by the architecture rather than the weights, show them in open systems and in Suno and Udio, and detect generated music from them at above 99 % accuracy in several scenarios. Large corpora of commercial output now exist - SONICS holds over 49,000 synthetic songs from Suno and Udio (Rahman et al., 2025) - but detection built on such traces is fragile under manipulation and on unseen generators (Afchar, Meseguer-Brocal, & Hennequin, 2025; Oh, 2026a), and benchmarks that compare detectors are themselves hard to keep consistent: an audit of ArtifactBench found three incompatible manifest lineages and audio that could no longer be fully reconstructed (Oh, 2026b).

2.3 The delivery codec as a confound

Lossy encoders remove the top of the spectrum by design, and the result looks like a model's band limit. The detector of Oh (2026a) initially labelled real music as generated after MP3 encoding, at a false-positive rate of 98.7 %, and had to be retrained across codecs. Reference collections are not exempt: MTG-Jamendo, a common source of human-made music in this field, distributes all its audio as 320 kbps MP3 (Bogdanov et al., 2019).

2.4 Measures

Spectral roll-off, the frequency below which a set share of a frame's energy lies, was introduced as a feature for telling speech from music (Scheirer & Slaney, 1997); it describes where the energy is, not where the spectrum ends. Spectral flatness, the ratio of the geometric to the arithmetic mean of the power spectrum, is 1 for white noise and falls towards 0 for a few strong partials (Gray & Markel, 1974). Inter-channel correlation measures how alike two channels are, and is a basic cue of perceived width (Faller & Baumgarte, 2003). The Fréchet Audio Distance compares the distribution of embeddings of generated audio with that of a reference set (Kilgour et al., 2019); its value depends on the embedding model and the reference, so scores computed with different embeddings or reference sets cannot be compared (Gui et al., 2024). On AIME's own human ratings, FAD computed with a music-trained CLAP embedding tracked human judgements of quality best of the metrics tested (Grötschla et al., 2025).

3.1 Data

AIME holds 6,000 tracks from twelve generators current in July 2024, and 500 human-made tracks from MTG-Jamendo, released for research with 15,600 human pairwise judgements (Grötschla et al., 2025); the generated audio is licensed CC BY 4.0. We drew 60 tracks per source - the first 60 in dataset order, taking each source's own shards first and filtering shared shards by source - for 780 tracks in all. No audio was kept after measurement.

Table 2 lists what the files are. Every stereo source is stored as 48 kHz WAV; MusicGen as 32 kHz mono; Riffusion as 44.1 kHz mono; AudioLDM 2 and Mustango as 16 kHz mono. AIME does not say how the commercial systems' audio was obtained or whether it passed through a lossy format on the way; the measurements below are of the files as published.

3.2 Processing

Each file was decoded at its own sample rate and channel count. From each, the 10-second window with the highest RMS level was taken - the rule AIME used for its listening clips. No loudness normalisation was applied, because every measure used is unchanged by gain. Nothing was resampled: all results are in absolute frequency, and a resampling filter short enough to be practical leaves images a few tens of decibels down, which a bandwidth measure reads as content the source never had. Short-time spectra used a Hann window of 2,048 points at 32, 44.1 and 48 kHz and 1,024 at 16 kHz (43–64 ms), with a hop of a quarter window; frames more than 60 dB below the loudest were dropped.

3.3 Measures

With X_t[k] the spectrum of frame t at bin k, f_k the bin frequency and K the number of bins up to the Nyquist frequency:

with γ = 0.85 and γ = 0.99, averaged over frames. The long-term average spectrum L[k], the mean of |X_t[k]|² over frames in decibels and smoothed over about 100 Hz, gives the bandwidth

with B₈₀ defined the same way. Where B₆₀ is not simply the file's Nyquist frequency, the steepness of the edge is the mean level 250–750 Hz below it minus the mean level 250–750 Hz above it, in dB per kHz; an edge of 40 dB per kHz or more is called steep. The share of energy above 16 kHz is the ratio of L summed above 16 kHz to L summed over all bins. Spectral flatness is

over the N bins from 50 Hz to 7.5 kHz, a band every source holds, per channel and then averaged. For the two channels x_L and x_R of a window of M samples, the zero-lag correlation is

computed over the whole window and over frames of 2,048 samples (43 ms at 48 kHz), and the change in level when the two are summed to mono is

which is 0 dB for identical channels, −3 dB for unrelated ones and falls without limit as they approach inversion. A file whose two channels were identical was treated as mono.

3.4 Validation

Before any real audio was measured, the pipeline was run on signals with known answers.

Two faults were found and fixed this way: a smoothing step that biased the top bins upwards, and a resampling step whose images read as bandwidth; the second is why nothing is resampled.

3.5 Statistics

Measures are reported as medians with interquartile ranges, since their distributions are skewed. Each generator was compared with the human reference by a two-sided Mann-Whitney U test, with the rank-biserial correlation r as effect size (positive when the generator's values are higher), and p-values Holm-corrected within each measure. α = .05. The human reference is itself lossy; section 5.4 returns to what that allows.

4.1 Where the spectrum ends

No source in the sample filled a full-band file (Figure 1, Table 3). The open systems ended at their decoding ceiling or just under it: AudioLDM 2 and Mustango at the 8.0 kHz Nyquist frequency of their 16 kHz files; MusicGen at a median of 15.3 kHz–15.4 kHz, under the 16 kHz of its 32 kHz codec; and Riffusion at 10.01 kHz (interquartile range 9.99 kHz–10.03 kHz), although its files run at 44.1 kHz - the maximum frequency of its spectrograms, not the rate of its files.

The commercial systems' files ended higher and more tightly: Stable Audio 1 at a median of 16.6 kHz (16.2 kHz–16.6 kHz), Suno v3 at 17.7 kHz, Suno v3.5 at 18.5 kHz, and Udio at 20.0 kHz. Stable Audio 2 ended lower and more variably (14.5 kHz; 10.3 kHz–16.3 kHz), as a gradual roll-off of its content rather than one edge. The human reference reached a median of 18.2 kHz (13.7 kHz–20.0 kHz).

4.2 The shape of the edge

Where the edge falls matters less than its shape. In 46 of 60 Stable Audio 1 tracks the spectrum ended in a steep edge, and those edges stood at one frequency - a median of 16.59 kHz. The same held for Suno v3 (26 tracks, at 18.84 kHz), Suno v3.5 (25, at 18.75 kHz) and Udio (34, at 20.02 kHz). Within each system these edges stood in a narrow band - the middle half of Stable Audio 1's spanned 94 Hz, Suno v3.5's 93 Hz and Udio's 64 Hz - across tracks of different genres. Music does not end at one frequency; a filter does. The human reference, known to be 320 kbps MP3, shows the same kind of edge in 15 tracks, at 20.04 kHz. Figure 3 sets these edges beside those LAME leaves on noise at 128, 192 and 320 kbps, which fall a few hundred hertz above each cluster. Steepness itself depends on the signal - music carries less energy just below its edge than noise does, so its edges measure less steep - and can flag a filter but not name an encoder.

A steep edge at one frequency is what a lossy encoder leaves, but also what a fixed low-pass stage inside a system would leave, so the data cannot say which produced these. They can say that the edges are not the Nyquist frequency of the files - none of the stereo files is Nyquist-limited - and not a property of the architectures as documented: Stable Audio decodes at 44.1 kHz (Evans, Carr, et al., 2024), yet its files end at 16.6 kHz.

4.3 The 85 % roll-off

The 85 % roll-off of every source lay between 0.57 kHz and 1.72 kHz (Table 3): most of a full mix's energy sits in the bass and low mids, so the frequency below which 85 % of it lies is low, whatever happens at the top. Even the 99 % roll-off stayed under 6 kHz for every source. The measure locates the body of the spectrum and cannot detect a cut at 16 kHz; values of 14–19 kHz reported as “85 % roll-off” for full mixes describe something else.

4.4 Flatness

Within the band every source holds, several generators were significantly flatter - more noise-like - than the human reference: Udio (median 0.038 against 0.006; p < .001, r = 0.57), Stable Audio 1 (p < .001, r = 0.47), Suno v3.5 (p = .008, r = 0.34), MusicGen Small and Large, AudioLDM 2 Large and Mustango. Suno v3 and Stable Audio 2 did not differ from it. Flatness also rises with genre and arrangement - a snare-heavy mix is flatter than a solo piano - and AIME's prompts differ in genre between sources, so these differences describe the samples; they are not a ranking of quality.

4.5 Stereo

Six sources are stereo in AIME: the human reference, Suno v3 and v3.5, Udio, and Stable Audio 1 and 2 (Figure 2, Table 4). None of the commercial systems showed the disruption the second hypothesis predicts. Suno v3.5 was the most strongly correlated source in the sample (median ρ = 0.93; 0.84–0.98), well above the human reference (ρ = 0.55; p < .001, r = 0.61); so were Udio (ρ = 0.83; p < .001) and Suno v3 (ρ = 0.84; p = .005). Their channels moved in opposite directions in a median of 0 % of frames, against 3.1 % for the human reference, and summing them to mono cost less: -0.17 dB for Suno v3.5 against -1.11 dB. The Stable Audio models did not differ from the human reference (ρ = 0.75 and 0.78; p = .142 and p = .077), though Stable Audio 2 held the most tracks with a negative overall correlation (7 of 60, against 4 for the human reference).

5.1 Paradigm does not set the band edge

The first hypothesis does not survive the comparison. An autoregressive system (MusicGen) ended at 15.4 kHz, and a diffusion system (AudioLDM 2) at 8.0 kHz; a diffusion system decoded at 44.1 kHz (Riffusion) ended at 10.0 kHz; and diffusion systems documented at 44.1 kHz (Stable Audio) delivered files that end at 16.6 kHz (Stable Audio 1) or lower. In every open system the edge is the ceiling of the representation the audio was decoded from - the codec's rate, the diffusion model's mel rate or the spectrogram's range (Table 1) - and not the paradigm. A codec's number of codebooks governs how much detail survives within its band (Défossez et al., 2023; Kumar et al., 2023), but the band itself is set by the rate the codec runs at, and codecs exist at 44.1 kHz.

5.2 The delivery codec comes before the model

The commercial systems' files carry steep edges at one frequency per system, of the kind a lossy encoder leaves, and the human reference - known to be MP3 - carries the same kind. Until the provenance of such files is known, a band limit measured in them cannot be attributed to the model. Three practices follow. Studies should obtain lossless exports where a service offers them and state the format they measured. They should report the steepness of the upper edge as well as its frequency, since a steep edge at one frequency points to a filter. And human references should be lossless, or matched in codec and bitrate to the material they are compared with.

5.3 The stereo image is narrow, not broken

The second hypothesis is not supported for the commercial systems measured. Suno and Udio delivered images more strongly correlated than the human reference, with less level lost on a mono sum; in production terms they are narrow and mono-safe, not phase-disrupted. The design the hypothesis rests on - channels coded separately - is documented for MusicGen's stereo variants (Meta AI, n.d.), which AIME does not contain; it remains a fair question for those models, not a finding about commercial music generation.

5.4 In post-production

For an engineer receiving generated audio, the evidence suggests an order of work. First, look at the top of the long-term spectrum: a steep edge at 16–20 kHz marks a band-limited file, and no equaliser will bring back what is not there. Content above the edge can only be synthesised - by an exciter that derives harmonics from the band below, or by neural bandwidth extension such as AudioSR, which extends inputs band-limited to between 2 and 16 kHz up to 24 kHz at a 48 kHz rate (H. Liu, Chen, et al., 2024) - and either adds plausible content, not the original. Second, read the correlation before touching the stereo image: the measured images call for widening, if anything, rather than repair, and the usual mastering check of mono compatibility applies as for any mix (Katz, 2015). Third, normalise loudness to the target of the destination, measured to ITU-R BS.1770-5: −23 LUFS for broadcast under EBU R 128, and the platform's own reference for streaming - Spotify, for one, normalises to −14 LUFS (European Broadcasting Union, 2023; International Telecommunication Union, 2023; Spotify, n.d.).

5.5 Limitations

• AIME was made in July 2024. Suno, Udio and Stable Audio have released later versions since, and Lyria and YuE are not in it; none of these results describes them.
• AIME does not state how the commercial audio was obtained, so a lossy stage cannot be confirmed or ruled out.
• The human reference is 320 kbps MP3; a lossless human reference would sharpen every comparison of bandwidth, and MP3's joint-stereo coding may also have changed how its channels relate.
• Sixty tracks per source, taken in dataset order, from 10-second windows; the sources also differ in genre and length, which move flatness and width.
• The measures are signal-level. They describe the files; they do not measure what listeners hear or prefer.
• FAD was not recomputed; this paper relies on the published evaluation of AIME for it (Grötschla et al., 2025).

Measured on public data, the common account of AI music's acoustic faults needs two corrections. The upper edge of a generated track is set by the representation it is decoded from and by the file it is delivered in, not by whether the model is autoregressive or diffusion; and the commercial systems measured deliver narrow, strongly correlated stereo, not disrupted phase. Many of their files end in steep edges between 16.6 and 20.0 kHz, each system's at one frequency, of the kind a lossy encoder or a fixed filter leaves. Deciding whether those edges belong to the models or to their delivery needs lossless exports and a lossless human reference, and the pipeline published with this paper is ready to measure them.

Every number in this paper comes from aime-measurements.csv (780 rows: one per track, identified by its AIME id, with every measure) or from the sources cited. The measures are computed by measure.py (Python, NumPy 2.4.6, SciPy 1.17.1). The audio itself is AIME's and is not redistributed here.