Pith. sign in

REVIEW 4 major objections 6 minor

Qwen-Music generates complete high-fidelity songs with vocals from text or a reference melody by separating semantic composition from acoustic rendering, and reports better results than leading commercial systems on most measured musicality

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 03:46 UTC pith:5MDKZOSV

load-bearing objection Solid systems report: Melody-CoT plus a careful tokenizer/render stack, with strong but closed-pipeline SOTA numbers against named commercial systems. the 4 major comments →

arxiv 2607.11699 v2 pith:5MDKZOSV submitted 2026-07-13 cs.SD

Qwen-Music Technical Report

Jin Xu , Kangdi Wang , Ruibin Yuan , Shun Lei , Xiong Wang , Xize Cheng , Xueyao Zhang , Yang Zhang
show 19 more authors
Yiheng Chen Yongqi Wang Yue Wang Zhifang Guo Zihan Liu Zijian Lin Dake Guo Hangrui Hu Lei Xie Linhan Ma Wei Xue Wenxiang Guo Xinfa Zhu Xipin Wei Yangze Li Yuanjun Lv Yuxuan Wang Yunfei Chu Zhiyong Wu
This is my paper
classification cs.SD
keywords music generationtext-to-musiccover song generationMelody-CoTmusic semantic tokensgenerative audio renderingpreference alignmentsinging voice
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This report introduces Qwen-Music, a system that writes full songs with singing either from text, lyrics, and musical tags, or as style-and-voice covers of a reference melody. Music is first compressed into a compact 25 Hz stream of single-codebook Music Semantic Tokens; a language model then plans and predicts those tokens, optionally first drafting a coarse melody plan (Melody-CoT); a generative renderer finally turns the tokens into 48 kHz stereo audio. The model is trained on more than five million hours of multilingual music with a quality-graded curriculum, then refined with supervised fine-tuning, offline preference optimization, and online policy optimization. Across 600 Chinese and English prompts it claims the best score on 13 of 16 objective musicality and audio-quality metrics, professional blind preference wins over several proprietary systems, and more accurate reference-melody preservation in cover generation. A reader who cares about controllable song creation would care because the work argues that long-range musical structure and waveform fidelity can be achieved together without forcing one model to do both jobs.

Core claim

Separating song generation into compact Music Semantic Tokens for autoregressive composition, an explicit melody-token chain-of-thought (Melody-CoT) that plans vocal contour before full-token generation, and generative stereo rendering from those tokens, when trained with a quality-graded curriculum and multi-stage preference alignment on over five million hours of data, produces songs that the authors measure as state-of-the-art on 13 of 16 objective metrics and preferred by professional raters over leading proprietary systems, while also cloning reference melodies more accurately for covers.

What carries the argument

Melody-CoT: before emitting full-mixture Music Semantic Tokens, the language model first produces a short sequence of relative-pitch melody tokens (from a downsampled vocal pitch contour) that sketches the vocal line. That intermediate plan structures original songs and, when taken from a reference, conditions covers so melody is preserved while style and voice remain controllable. The surrounding stack is a 25 Hz single-codebook tokenizer, the autoregressive LLM, and a DiT-plus-Spec-VAE Band-Mode Refiner renderer.

Load-bearing premise

The claim that Qwen-Music is better overall depends on automatic musicality scores and an internal preference predictor built and judged largely inside the authors’ own evaluation pipeline; if those scores do not track what independent listeners prefer, the reported wins do not hold.

What would settle it

Run a fresh blind A/B preference study on an independently curated set of bilingual prompts (new lyrics, tags, and genres), with raters who have never evaluated this system, pitting Qwen-Music against Suno V5.5 and MiniMax Music 2.6 under identical inputs; if preference flips or SongBench and SongEval rankings reverse substantially, the central superiority claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Open text-to-song generation with control over genre, mood, instruments, and vocal attributes can match or beat closed commercial systems on measured musicality and preference.
  • Cover song systems can keep a reference melody more tightly while still following new style and voice instructions via explicit melody-token conditioning.
  • Quality-bucket pre-training followed by supervised start, offline DPO, and online GSPO is a workable path to raise musicality and instruction following at multi-million-hour scale.
  • Low-bitrate semantic tokens plus a separate generative stereo renderer can recover high-fidelity 48 kHz audio without making the language model model waveforms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Explicit intermediate melody planning is likely useful beyond songs wherever long creative sequences need structure before detail (film cues, multi-track arrangement, lyric-timed scoring).
  • The section-level versus unique-section Melody-CoT trade-off (tighter cloning versus freer style control) suggests product systems will need both conditioning modes exposed to users.
  • If the internal quality predictor generalizes, genre-normalized quality curricula may become standard for any large generative audio model trained on messy web-scale music.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. Qwen-Music is a large-scale song generation system for text-to-music and reference-melody cover generation. It separates semantic composition from acoustic rendering via three components: a 25 Hz single-codebook Music Semantic Tokenizer (four-stage Conformer+VQ recipe), an autoregressive Qwen-Music-LLM with Melody-CoT (section-level and unique-section-level melody planning/conditioning), and Qwen-Music-Render (semantic-conditioned DiT, Spec-VAE, Band-Mode Refiner) for 48 kHz stereo. The LLM is trained on >5M hours with a quality-graded curriculum (Q1–Q6) and multi-stage post-training (SFT, offline DPO, online GSPO). On 600 bilingual prompts the authors report best scores on 13/16 SongBench/SongEval/AudioBox-Aesthetic metrics, professional A/B preference wins or ties versus MiniMax, Mureka, and Suno, and stronger Melody MAE on cover tasks than several proprietary systems.

Significance. If the reported rankings hold under independent scrutiny, this is a substantial systems contribution to full-song generation with vocals: a unified text-to-music and cover pipeline, an explicit melody intermediate (Melody-CoT), a carefully engineered low-bitrate semantic tokenizer, and a high-fidelity stereo renderer with documented acoustic filtering and ablations. The evaluation package is unusually broad for an industry technical report—objective multi-suite tables, 50 professional raters with genre-wise Bradley–Terry, an external Artificial Analysis leaderboard entry, and renderer reconstruction benchmarks—and the architecture is described at a level that is useful to the field even without full model release. The main scientific value is the composition/rendering split plus melody-as-CoT as a controllable interface, not a new theoretical result.

major comments (4)
  1. [Table 5, §4.1.1, Figure 2] Table 5 and §4.1.1: The central “13 of 16” SOTA claim is load-bearing but reported as point estimates only. There are no confidence intervals, bootstrap/paired tests, or prompt-level variance for SongBench/SongEval/AudioBox scores, nor for the A/B win rates in Figure 2 (despite three raters per pair). Several margins are small (e.g., SongEval Memorability/Naturalness ties or near-ties with Mureka; AudioBox production metrics where competitors win). Without inferential statistics or released per-prompt scores, the ranking cannot be assessed for stability on this closed 600-prompt set.
  2. [§2.3, Abstract, Tables 6–7] §2.3 and Abstract: Melody-CoT is presented as a key novelty that improves creativity, musicality, and structural coherence for text-to-music, yet the only controlled comparisons of section vs unique-section conditioning appear in cover-song Tables 6–7. There is no ablation of text-to-music with vs without Melody-CoT (or vs direct full-token generation) on SongBench/SongEval or human preference. The text-to-music benefit of the mechanism is therefore asserted rather than measured, which weakens the architectural claim relative to the cover-song results.
  3. [§4.1.2, Eq. (4), Tables 6–7] §4.1.2, Eq. (4), Tables 6–7: Reference-melody preservation is quantified by Melody MAE after relative-MIDI conversion and DTW using the authors’ own melody tokenizer pipeline. This metric is not validated against human melody-similarity judgments, and Suno comparisons are restricted to AI-generated references because Suno does not accept real-world songs. On the AI-generated set, proprietary systems still lead most SongBench and tag-following dimensions while Qwen-Music leads Melody MAE—so the abstract’s “preserves reference melodies more accurately” claim is metric-specific and should be scoped, with human melody-preservation ratings or an independent pitch tracker as a check.
  4. [§3.1–3.2, §4.1] §3.1–3.2 and §4.1: Quality bucketing, DPO, and GSPO rely on an internal MOS musicality predictor trained on professional annotations, while tag following uses Gemini 3.1 Pro and intelligibility uses Qwen3-ASR PER. Final claims also cite external proprietary A/B and Artificial Analysis, which partially mitigates circularity, but the manuscript does not report independence checks (e.g., correlation of internal MOS with the professional A/B outcomes on the same clips; hold-out prompts outside the reward-data regime; third-party re-scoring). For a journal-level SOTA claim, either release evaluation audio/prompts or add an independent human study that does not reuse the training reward stack.
minor comments (6)
  1. [Abstract] Abstract grammar: “which create entirely new songs” → “which creates”; “Cover Song Generation, which reinterprets” is fine but parallel structure with the first clause should be cleaned.
  2. [§2.2.4] §2.2.4 heading typography: “Fine-T uning” appears with a broken space/hyphen.
  3. [Figure 6] Figure 6 label “No-NeqSpec” / “LostSpec” is unexplained in the caption; define band names and the Nyquist path briefly.
  4. [Table 4] Table 4: Bradley–Terry ratings are within-genre only (as noted), but the large spread (e.g., Jazz & Blues Mureka 202 vs others ~900+) would benefit from pair counts and uncertainty so readers can judge sparsity.
  5. [Table 5] §4.1.1: PER is reported after ×100; state units explicitly in the table header (e.g., PER %) to avoid confusion with raw rates.
  6. [§4] Clarify whether the 600-prompt set, rewritten tags/lyrics, and cover reference sets will be released; even without model weights, a fixed public eval set would strengthen the report.

Circularity Check

1 steps flagged

No derivation-by-construction circularity: Qwen-Music is an empirical systems report whose SOTA and preference claims rest on external A/B tests, proprietary baselines, and multi-suite metrics rather than on equations that reduce to fitted inputs.

specific steps
  1. other [Sec. 3.1–3.2 (internal MOS); SongBench citation Wu et al. 2026; Sec. 4.1 Melody MAE / Eq. 4]
    "Following the multi-aspect song-quality assessment protocols of SongBench (Wu et al., 2026) and SongEval (Yao et al., 2025), we train an internal MOS-based reward model. ... we further measure reference-melody preservation using Melody MAE. Specifically, we convert both the generated cover and the reference song into relative MIDI representations following Eq. 4"

    Not true circularity: the internal MOS shapes training data and preference optimization, and SongBench has author overlap, while Melody MAE reuses the conditioning representation. These create mild self-reinforcement risk, but the headline SOTA/preference claims are still empirical comparisons to external proprietary systems and professional A/B judges, not quantities forced equal to the training reward or to Eq. 4 by definition.

full rationale

This manuscript is an engineering/systems technical report, not a first-principles derivation paper. There is no claimed theoretical prediction chain in which a quantity is fitted and then re-presented as an independent forecast, nor any uniqueness theorem imported from overlapping authors that forces the architecture. The Melody-CoT, tokenizer stages, DiT/Spec-VAE render, quality-graded curriculum, and DPO/GSPO pipeline are design and training choices evaluated empirically. Final claims compare Qwen-Music to Suno, MiniMax, and Mureka under fixed prompts, plus professional blind A/B votes and an external Artificial Analysis leaderboard entry—evidence outside the training reward. Mild author overlap on SongBench (Shun Lei) and use of an internal MOS reward for curriculum/post-training introduce possible evaluation bias, but do not make the 13/16 SOTA count or preference rates true by construction: those scores are still measured outputs of generated audio against independent systems and human raters. Melody MAE uses the same relative-MIDI map as conditioning (Eq. 4), which aligns metric with method, yet all systems are scored identically, so the ranking is comparative rather than definitional. Score 1 reflects only that minor self-citation/metric-alignment residue, not load-bearing circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

As a large-scale empirical systems paper the central performance claims rest on architectural choices, training curricula, reward models, and evaluation protocols rather than mathematical axioms. Free parameters dominate; invented entities are the named modules of the system itself.

free parameters (5)
  • VQ codebook size and rate
    32768-entry single codebook at 25 Hz (375 bit/s) chosen by design; utilization and bitrate are engineering knobs that directly affect LLM sequence length and renderer fidelity.
  • Melody downsampling and relative-MIDI range
    50 Hz RMVPE → 8× median pool → 6.25 Hz, clipped relative MIDI offsets mapped to 256 tokens; the factor-8 and clip range are hand-chosen.
  • Quality bucket percentiles (Q1–Q7)
    Genre-normalized MOS ranking cut at 90/75/50/25/5/1 %; the thresholds and the decision to drop Q7 control the entire pre-training curriculum.
  • DPO/GSPO reward weights and iteration counts
    Musicality MOS + instruction-following rewards from Qwen3.5-Omni/Qwen3-ASR; relative weighting and number of offline/online iterations are free optimization choices.
  • DiT/Spec-VAE latent dimension and CFG strategy
    128-dim latents, text-drop CFG, cross-attention lengths (256/1536), and Band-Mode frequency splits are design parameters that determine reported audio quality.
axioms (4)
  • domain assumption Discrete 25 Hz Music Semantic Tokens retain sufficient semantic and melodic information for an autoregressive LLM to compose coherent full songs.
    Stated as design principle in §2.1–2.2; underpins the entire composition–render split.
  • domain assumption Relative-MIDI melody tokens extracted by RMVPE + median pooling capture the compositional contour while discarding absolute pitch and timbre.
    §2.3.1; required for Melody-CoT and cover-song conditioning to work as claimed.
  • ad hoc to paper Internal MOS reward model trained on professional annotations is a reliable proxy for human musical preference.
    §3.1; used both for data curriculum and for DPO/GSPO preference pairs.
  • domain assumption Automatic metrics (SongBench, SongEval, AudioBox-Aesthetic, Gemini tag scores, Qwen3-ASR PER, Melody MAE) correlate sufficiently with expert listening preference.
    §4.1; load-bearing for the 13/16 SOTA and cover-song claims.
invented entities (3)
  • Music Semantic Tokens (25 Hz single-codebook) no independent evidence
    purpose: Compact discrete interface between tokenizer and LLM that preserves melody/semantics for long-horizon composition.
    Defined by the four-stage tokenizer recipe; no independent existence outside this system.
  • Melody-CoT (section-level and unique-section-level) no independent evidence
    purpose: Explicit intermediate melody plan that improves structural coherence and enables reference-melody cloning.
    Core claimed novelty of the LLM stage; validated only by the paper’s own ablations and cover metrics.
  • Spec-SnakeBeta + Band-Mode Refiner no independent evidence
    purpose: Frequency-aware activation and band-specific magnitude/phase residual correction for high-fidelity spectrogram reconstruction.
    Architectural inventions of the renderer; gains shown only on the authors’ reconstruction suite.

pith-pipeline@v1.1.0-grok45 · 27407 in / 3307 out tokens · 28688 ms · 2026-07-14T03:46:02.645978+00:00 · methodology

0 comments
read the original abstract

In this report, we introduce Qwen-Music, a powerful music generation model capable of producing highly musical and high-fidelity songs with complete vocal singing. Qwen-Music supports two core tasks: Text to Music Generation, which create entirely new songs from text descriptions, lyrics, and musical attributes, and Cover Song Generation, which reinterprets existing songs with different styles and vocal characteristics. Architecturally, Qwen-Music integrates three core components: Qwen-Music-Tokenizer, Qwen-Music-LLM, and Qwen-Music-Render. Qwen-Music-Tokenizer compresses audio into a 25 Hz single-codebook stream of Music Semantic Tokens that preserve semantic and melodic information for LLM prediction. Based on these tokens, Qwen-Music-LLM performs autoregressive music semantic modeling, with a key novelty being a melody-token-based chain-of-thought (Melody-CoT) mechanism that plans melodies before full-song generation, improving creativity, musicality, structural coherence, and reference-audio-based melody cloning. To overcome the fidelity limitations of discrete semantic tokens, Qwen-Music-Render performs generative stereo rendering, enriching acoustic details and producing high-fidelity stereo waveforms. Finally, we train Qwen-Music-LLM on more than 5 million hours of multilingual music data covering hundreds of languages. We first apply quality-aware pre-training curriculum, then use progressive post-training, comprising supervised initialization, offline DPO, and online GSPO, to further improve musicality and instruction-following ability. Across 600 Chinese and English prompts, Qwen-Music achieves state-of-the-art results in 13 of 16 objective musicality and audio-quality metrics. Professional evaluators also prefer Qwen-Music over leading proprietary systems. For cover song generation, Qwen-Music preserves reference melodies more accurately than leading proprietary systems.

Figures

Figures reproduced from arXiv: 2607.11699 by Dake Guo, Hangrui Hu, Jin Xu, Kangdi Wang, Lei Xie, Linhan Ma, Ruibin Yuan, Shun Lei, Wei Xue, Wenxiang Guo, Xinfa Zhu, Xiong Wang, Xipin Wei, Xize Cheng, Xueyao Zhang, Yangze Li, Yang Zhang, Yiheng Chen, Yongqi Wang, Yuanjun Lv, Yue Wang, Yunfei Chu, Yuxuan Wang, Zhifang Guo, Zhiyong Wu, Zihan Liu, Zijian Lin.

Figure 1
Figure 1. Figure 1: Qwen-Music is a powerful and controllable music generation model capable of producing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Blind A/B preference results between Qwen-Music and leading proprietary systems, including [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: External leaderboard result on the Artificial Analysis Music with Vocals Leaderboard. Qwen [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the Qwen-Music inference pipeline. Given a user request, Qwen-Music first [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of Qwen-Music-Tokenizer. The tokenizer maps music waveforms to 25 Hz Music [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of Qwen-Music-Render. Music Semantic Tokens and text conditions guide a semantic [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Spec-SnakeBeta analysis. (a) Activation curves: the standard SnakeBeta (black dashed) applies a single shared α to all frequency bins; Spec-SnakeBeta adapts the periodic modulation per frequency, pro￾ducing distinct nonlinear shapes at 0.5, 3, 10, and 20 kHz. (b) Deviation of learned αf from log-frequency initialization (∆αf = αlearned − αinit) across encoder layers—positive values indicate the network inc… view at source ↗
Figure 8
Figure 8. Figure 8: High-frequency cutoff detector examples. The blue curve is the Welch spectrum of a 10 s [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.