USESUNO RESEARCH TECHNICAL NOTE 001 · SEPTEMBER 2026

Suno Audio Conversion Baseline: Source, MP3 and PCM WAV

Technical measurements from 48 controlled Suno v6 tracks

UseSuno ResearchIndependent research by UseSuno

48 tracks · 24 requests · 12 matched pairs · Suno v6 · 48/48 processed · 0 errors

Abstract

Before comparing audio measurements across Suno generations or model versions, we established a conversion baseline using 48 controlled Suno v6 tracks. Each decoded public source was compared with a 192 kbps CBR MP3 conversion and a local signed 16-bit PCM WAV representation. For five of six metrics, median variation between the two outputs of the same generation request exceeded the corresponding MP3 conversion effect; the ratios were 150.7× for duration and 108.9× for low-frequency energy. File size was the exception. MP3/source RMS ratios were below 1.0 in all 48 tracks (median 0.969979) and were similar in separate synthetic checks through the bundled WASM and FFmpeg libmp3lame paths. Local PCM WAV closely reproduced decoded-source RMS (median ratio 0.99998468) while increasing file size. These results define a sample- and pipeline-specific measurement baseline, not a perceptual quality assessment.

The findings apply to this controlled v6 sample and the specified conversion paths; they do not establish universal format behavior or audible differences.

Keywords: Suno · audio conversion · generation variability · MP3 · PCM WAV · measurement baseline

Key Findings

For five of six metrics, median V1/V2 generation variation exceeded the corresponding MP3 conversion effect. The ratios ranged from 1.36× for RMS to 150.7× for duration; source file size was the exception at 0.085×.
The tested MP3 pipeline produced a consistent RMS signature: all 48 MP3/source ratios were below 1.0, with a median of 0.969979. Separate synthetic checks through the bundled WASM and FFmpeg libmp3lame paths produced ratios near 0.970.
Local 16-bit PCM WAV closely reproduced decoded-source RMS (median ratio 0.99998468). Samples above 1.0 were clamped before PCM quantization; this does not establish whether the source was hard-clipped.
File-size ratios describe representation in this sample: median 1.396587× for MP3 and 11.170870× for local PCM WAV versus source. They are not measures of audio quality.
The bass-focused manipulation was weak and mixed: 7/12 matched pairs moved in the intended direction and 5/12 did not. No strong bass-specific causal conclusion is drawn.

1. Introduction

Format comparisons can be difficult to interpret when independently generated tracks differ before conversion. This technical note establishes a baseline for the measurement pipeline first, then compares conversion effects on each decoded source with variation between two outputs from the same Suno generation request.

The central question is methodological: how large are the measured conversion effects relative to generation-to-generation variation in this controlled v6 sample? The results concern file and decoded-signal measurements, not perceived sound quality.

2. Study Design

The formal study used 24 generation requests with two outputs per request, forming 12 matched bass-focused and reference pairs. The same decoded public source for each of the 48 outputs was measured as source audio, converted to 192 kbps CBR MP3, and rendered locally as signed 16-bit PCM WAV.

Within each request, V1 and V2 are separate Suno outputs. Vocal matched pairs used the same lyrics. The queue declared model v6. Processing completed serially for all tracks, with no errors.

Table 1. Study design
Measure	Study specification
Model	v6
Generation requests	24
Tracks	48
Matched pairs	12
Vocal tracks	16
Instrumental tracks	32
MP3	192 kbps CBR · WASM / LAME-family path
WAV	Local signed 16-bit PCM
Processing success	48 / 48
Errors	0

The earlier 16-track pilot is kept separate and is not pooled into any formal-study statistic. Per-track and aggregate measurements use anonymous sample and pair labels; public song and account identifiers, internal URLs, local paths, and processing logs are omitted from the released extracts.

3. Results

3.1 Conversion Baseline

For context, Table 2 summarizes the median file-size and RMS ratios for each representation. The 48-track MP3/source size ratio was 1.396587× (range 1.243162–1.462192×); every converted MP3 was larger than its corresponding source. Local PCM WAV was 7.998543× the MP3 size at the median. These are characteristics of the tested conversions, not indicators of audio quality or recovered information.

Table 2. Median size and RMS ratios relative to decoded source
Representation	Size ratio	RMS ratio	Interpretation
Decoded source	1.000000×	1.000000	Reference
192 kbps CBR MP3	1.396587×	0.969979	Tested encode/decode path
Local 16-bit PCM WAV	11.170870×	0.99998468	Decoded source rendered as PCM

3.2 Generation Variation Relative to Conversion

For five of six compared metrics, the median difference between V1 and V2 outputs from the same request exceeded the median MP3 conversion effect measured from a source. Generation variation was 150.7× the conversion effect for duration and 108.9× for low-frequency energy; source file size was the exception. These descriptive ratios apply only to this sample.

Log-scale ratios of generation variation to MP3 conversion effect: duration 150.7, low-frequency energy 108.9, crest factor 10.31, peak 2.13, RMS 1.36, and source file size 0.085. The vertical 1.0 line marks equal effects. — Figure 1. Generation variation relative to the MP3 conversion effect across six metrics on a logarithmic scale. The vertical 1.0 reference marks equal effects; values above it indicate larger V1/V2 variation.

Table 3. Median generation difference and MP3 conversion effect
Metric	V1/V2 difference	MP3 effect	Ratio
Duration	5.2 s	0.0345 s	150.7×
Low-frequency energy ratio	0.035522	0.000326	108.9×
Crest factor	0.512733	0.049727	10.31×
Peak amplitude	0.062548	0.029337	2.13×
RMS	0.007281	0.005373	1.36×
Source file size	105,500 bytes	1,239,000 bytes	0.085×

For source file size, the median V1/V2 difference was 105,500 bytes, compared with a 1,239,000-byte median MP3 conversion effect.

3.3 MP3 Pipeline RMS Signature

The median MP3/source RMS ratio was 0.969979 (mean 0.970033; range 0.969773–0.970654). The ratio was below 1.0 for all 48 tracks in the defined 192 kbps encode/decode pipeline.

Jittered dot plot of all 48 MP3-to-source RMS ratios on an enlarged observation scale, with a clearly marked scale break before the 1.000 source reference. Every ratio is below 1.0; median 0.969979. — Figure 2. MP3/source RMS ratios for the 48 formal-study tracks. The observation range is enlarged, with an axis break before the 1.000 source reference. All ratios were below 1.0; the median was 0.969979. RMS is an amplitude measurement, not a perceptual quality score.

In a separate synthetic check, a 30-second, 48 kHz stereo 440 Hz sine wave at 0.5 peak produced a central RMS ratio of 0.970046 through the bundled WASM path and 0.970060 through FFmpeg libmp3lame. The check does not identify whether the level difference comes from encoding, MP3 coding and decoding, or their interaction.

3.4 Local PCM WAV Reproduction of Decoded RMS

Local signed 16-bit PCM WAV had a median WAV/source RMS ratio of 0.99998468 across the 48 tracks (range 0.99998408–0.99998488). Figure 3 shows each track relative to the decoded-source reference. This is a local conversion, not an official Suno WAV export.

Scatter plot of 48 local WAV RMS differences in dataset order, with a zoomed vertical scale from minus 16 to minus 15 ppm. Median minus 15.3 ppm. Equal RMS is zero ppm, outside the displayed range. — Figure 3. Track-level deviation of local PCM WAV/source RMS from 1.0, expressed in parts per million. Each point represents one track in dataset order. The vertical axis is zoomed to −16 to −15 ppm; equal RMS (0 ppm) is outside this range. The observed range was −15.9 to −15.1 ppm, with a median ratio of 0.99998468 (−15.3 ppm, dashed line).

File size increased by a median 11.170870× versus source and 7.998543× versus MP3. The larger PCM representation does not restore information absent from decoded source audio. Source samples above 1.0 were clamped to [-1, 1] before 16-bit PCM quantization; the decoded values alone do not establish hard clipping because the source quantization rail is unknown (Table 4).

3.5 Why the Bass Hypothesis Remained Unresolved

The matched-pair prompt manipulation was weak and mixed. Seven of 12 pairs moved in the intended direction, five moved in the opposite direction, and the median paired difference was +0.011130.

This study does not establish that bass-heavy Suno audio is more prone to clipping or MP3 degradation. Testing reports specific to v5.5 would require a separate historical v5.5 sample; the present measurements are from v6.

Zero-centered dot plot of 12 matched-pair differences in source low-frequency energy ratio. Seven pair values are positive and five negative; the median difference is plus 0.011130. — Figure 4. Matched-pair mean difference in source low-frequency ratio, bass-focused minus reference. Seven of 12 pairs were positive and five were negative; the median difference was +0.011130.

3.6 Full-Scale Decoded Samples

Decoded floating-point source samples exceeded 1.0 in 12 of 48 tracks, with 36 above-full-scale samples in total. The maximum decoded sample amplitude was 1.092280. These decoded values alone do not prove hard clipping because the source quantization rail is unknown. Local WAV conversion clamps samples to [-1, 1] before 16-bit PCM quantization; its maximum decoded sample was 1.000000.

Table 4. Full-scale decoded sample observations
Observation	Result
Tracks with samples ≥ 0.99	13 / 48
Tracks with samples > 1.0	12 / 48
Above-full-scale decoded samples	36
Maximum decoded source sample	1.092280
Local WAV samples > 1.0 after clamp	0 / 48 · maximum 1.000000

4. Practical Interpretation

4.1 Same-source format comparisons

To estimate a format-conversion effect, compare the same decoded source across formats. Comparing one track’s source with another track’s MP3 would mix conversion differences with generation-to-generation variation.

4.2 Generation and model comparisons

Comparisons across prompts, generations, or model versions need repeated outputs and a matched measurement path. The V1/V2 differences in this sample often exceeded the MP3 effect, so a single output per condition could obscure smaller conversion changes.

4.3 File representation

MP3 and local PCM WAV file sizes describe the tested representations and conversion settings. Larger files do not establish better audio quality, and a local PCM rendering cannot restore information absent from decoded source audio.

5. Conclusion

The main result of this baseline study is methodological: format-conversion effects and generation-to-generation effects should not be conflated. For five of six compared metrics, V1/V2 variation exceeded the MP3 conversion effect in this controlled v6 sample, including 150.7× for duration and 108.9× for low-frequency energy.

The tested MP3 pipeline produced a consistent RMS ratio near 0.970, while local PCM WAV closely reproduced decoded-source RMS after clamping samples to the signed 16-bit range. File-size differences characterize these representations; they are not quality measurements. The bass-focused result remains inconclusive.

These sample-specific baselines can inform future comparisons of Suno generations, model versions, and reported audio anomalies. Such work should compare formats from the same source and account for variation across repeated generations.

6. Limitations

The sample does not represent all Suno songs, prompts, users, models, or future generations.
The study does not establish audible quality differences or listener preference.
This is not a comparison of official Suno WAV and official Suno MP3. The WAV is a local 16-bit PCM conversion of decoded source audio.
File sizes do not reveal the source codec, bitrate, bit depth, or lossless/lossy status.
The bass manipulation was weak and mixed, so it does not establish a bass-specific causal effect.

All measurements describe this controlled v6 dataset and the defined conversion pipeline.

7. Data Availability

The sanitized measurements and analysis artifacts used in this report are available below. The per-track extract uses anonymous sample and pair labels; public song and account identifiers, internal URLs, local paths, and processing logs are omitted.

Primary statistics use only the formal 48-track study. No reproduction repository is published for this runner.

How to cite this report

UseSuno. (2026). Suno Audio Conversion Baseline: Source, MP3 and PCM WAV. UseSuno Research. https://usesuno.com/research/suno-audio-download-report/