MuLaCover: Flexible Cover Song Generation via Symbolic Tonal Control

Lynn Yi; Zihan Xiong; Junjie Cao; Xiaolong Weng; Yingzhe Ma; Gus Xia; Jinliang Liu; Yuxin Xie

cs.SDOct 19, 2026

MuLaCover: Flexible Cover Song Generation via Symbolic Tonal Control

Authors: Lynn Yi, Zihan Xiong, Junjie Cao, Xiaolong Weng, Yingzhe Ma, Gus Xia, Jinliang Liu, Yuxin Xie,

Organizations: MBZUAI · Peking University · MuLa Labs, Vera Praxis Lab · Nanjing University of the Arts ·

Abstract

A cover song re-renders an existing piece, most commonly by preserving its tonal content -- melody and chord progression -- while reshaping other musical attributes. This suggests a natural formulation for generative cover synthesis with pretrained music language models: explicitly control tonal content, while specifying lyrics and style through text. Existing approaches condition on frame-level features, tightly coupling generation to the reference's timing and structure, and limiting flexibility such as tempo variation, structural rearrangement, or partial conditioning. To this end, we present MuLaCover, a cover generation model that conditions a pretrained text-to-song foundation model on an editable symbolic lead sheet. Our design enables alignment-free control: generated outputs preserve the tonal signature of the source without frame-level alignment, allowing flexible timing, structure, and segment-level generation. We further introduce a training curriculum with partially mismatched audio-symbolic pairs, improving diversity and robustness. We evaluate MuLaCover with objective metrics, standard subjective ratings, and in-depth expert interviews that provide fine-grained diagnostic insights. Experiments show that MuLaCover achieves state-of-the-art performance among open-source systems and is competitive with leading commercial models such as Suno-v5.5. Audio samples and code are available at our project page: https://mulalabs.ai/mulacover/

Figures & tables

Figure 1: Overview of MuLaCover . (a) The pretrained text-to-song backbone has

N

transformer layers consuming lyrics

\ell

, style

s

, and audio tokens

a

; at its last

L

layers, the Gated Adaptive Cross-Attention (GACA) module injects the symbolic tonal-content condition

c

into the hidden states

h_{l}

. (b) Inside a GACA block,

c

is refined by a layer-specific self-attention (SA) into

c_{l}

, then merged into

h_{l}

through a cross-attention (CA) update modulated by a learned,

\tanh

-bounded gate

g_{l}

. (c) Cross-cover strategy in Stage 3: colors denote different renditions of the same song. Stages 1–2 draw all four signals

\ell,s,a,c

from the same rendition; Stage 3 draws

c

from a different rendition than

\ell,s,a

. (d) Partial conditioning via attention masking on

c

\rho{=}0

keeps the interval

[\alpha T,\beta T]

( outpainting ),

\rho{=}1

keeps its complement ( inpainting ); the default

(\alpha,\beta,\rho){=}(0,1,0)

recovers the full-song setting.

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
SongEcho [ 11 ]	7.88	6.11	7.26	7.54	3.70	3.57	3.37	3.47	3.46	0.26	0.38	0.39	0.88	1.55
ACE-Step-1.5 [ 12 ]	8.29	6.30	7.68	7.86	4.48	4.46	4.26	4.35	4.36	0.21	0.30	0.40	0.95	1.10
Suno-v5.5 [ 53 ]	8.30	6.44	7.66	7.92	4.61	4.60	4.42	4.52	4.51	0.11	0.50	0.33	0.98	1.11
MuLaCover	8.34	6.34	7.70	7.93	4.59	4.59	4.34	4.44	4.47	0.13	0.41	0.48	0.97	1.05

Table 1: Objective evaluation on the full-song cover task (Task 1).

\uparrow

\downarrow

indicate higher / lower is better. Bold marks the best and underline the second-best result in each column.

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
SongEcho [ 11 ]	7.58	5.59	6.80	7.44	3.14	3.04	2.85	2.87	2.92	0.55	0.25	0.63	0.41	5.94
ACE-Step-1.5 [ 12 ]	8.26	5.91	7.66	7.84	4.22	4.20	3.95	4.08	4.08	0.36	0.34	0.50	0.52	4.92
MuLaCover	8.33	6.31	7.69	7.92	4.57	4.57	4.32	4.42	4.45	0.12	0.42	0.50	0.70	2.60

Table 2: Objective evaluation on the partial-reference cover task (Task 2), averaged over the three masking configurations (continuation, inpainting, outpainting).

\uparrow

\downarrow

indicate higher / lower is better. Bold marks the best.

Method	Identity Preserveness ( $\uparrow$ )	Style Controllability ( $\uparrow$ )	Creativeness ( $\uparrow$ )	Overall Musicality ( $\uparrow$ )
SongEcho [ 11 ]	$2.84\pm 0.19$	$2.84\pm 0.23$	$2.56\pm 0.21$	$2.28\pm 0.21$
ACE-Step-1.5 [ 12 ]	$3.53\pm 0.21$	$2.93\pm 0.23$	$3.03\pm 0.21$	$3.31\pm 0.22$
Suno-v5.5 [ 53 ]	$\mathbf{3.81\pm 0.19}$	$\mathbf{3.72\pm 0.21}$	$3.51\pm 0.20$	$\mathbf{3.88\pm 0.21}$
MuLaCover	$3.45\pm 0.20$	$3.71\pm 0.25$	$\mathbf{3.65\pm 0.22}$	$3.75\pm 0.22$

Table 3: Listening study: mean

\pm

95% confidence interval on each metric (5-point Likert; higher is better). Bold marks the best and underline the second-best result in each column.

Training schedule
	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Variant	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
MuLaCover -small	8.28	6.36	7.67	7.89	4.58	4.56	4.33	4.42	4.45	0.14	0.38	0.49	0.98	1.04
w/o Warmup	8.07	6.13	7.44	7.72	4.41	4.40	4.13	4.24	4.23	0.16	0.40	0.54	0.11	9.75
w/ SP	8.24	6.43	7.63	7.87	4.57	4.55	4.33	4.41	4.44	0.16	0.35	0.45	0.98	1.04
Condition signal

Table 4: Ablation studies on the full-song cover task. The first row is the full MuLaCover ; rows below ablate one component at a time, grouped by training schedule , condition signal , and architecture .

\uparrow

\downarrow

indicate higher / lower is better.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Stage 1	Stage 2 (w/o mask)	Stage 2 (w/ mask)	Stage 3
Learning rate	$2{\times}10^{-4}$	$2{\times}10^{-4}$	$2{\times}10^{-4}$	$2{\times}10^{-5}$
Epochs	3	2	2	2
Avg. bs/GPU	9	1	1	1
$(\alpha,\beta,\rho)$	$(0.0,1.0,0.0)$	$(0.0,1.0,0.0)$	$\star$	$\star$
Trainable modules	GACA	GACA	GACA	GACA & LoRA
Trainable params	${\sim}300$ M	${\sim}300$ M	${\sim}300$ M	${\sim}400$ M

Appendix

Table 5: Per-stage training hyperparameters. GACA: gated adaptive cross-attention. Entries marked

\star

follow the dynamic mask schedule described below.

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
ACE-Step-1.5 $\lambda=0.1$	8.25	5.58	7.64	7.80	4.21	4.17	3.93	4.08	4.07	0.37	0.39	0.56	0.08	12.40
ACE-Step-1.5 $\lambda=0.2$	8.23	5.87	7.64	7.78	4.28	4.25	4.03	4.14	4.17	0.29	0.37	0.50	0.25	8.03
ACE-Step-1.5 $\lambda=0.3$	8.28	6.30	7.67	7.86	4.48	4.46	4.25	4.34	4.36	0.21	0.30	0.40	0.95	1.10
ACE-Step-1.5 $\lambda=0.4$	8.28	6.40	7.64	7.87	4.51	4.49	4.29	4.37	4.38	0.13	0.27	0.36	1.00	1.00
ACE-Step-1.5 $\lambda=0.5$	8.27	6.43	7.61	7.86	4.51	4.50	4.29	4.38	4.38	0.11	0.27	0.34	1.00	1.00

Appendix

Table 6: Objective evaluation results under different audio cover control strengths

\lambda

for ACE-Step 1.5.

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
SongEcho [ 11 ]	7.49	5.44	6.63	7.43	3.06	2.98	2.76	2.77	2.83	0.64	0.19	0.70	0.25	7.59
ACE-Step-1.5 [ 12 ]	8.26	5.86	7.65	7.83	4.19	4.17	3.91	4.05	4.04	0.36	0.35	0.51	0.38	6.18
MuLaCover	8.33	6.33	7.71	7.93	4.59	4.59	4.33	4.44	4.47	0.11	0.42	0.51	0.63	3.15

Appendix

Table 7: Partial-reference cover, continuation configuration. Same metrics as Table 1 .

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
SongEcho [ 11 ]	7.41	5.43	6.56	7.32	2.87	2.76	2.63	2.58	2.67	0.64	0.20	0.72	0.25	7.61
ACE-Step-1.5 [ 12 ]	8.23	5.71	7.62	7.82	4.17	4.13	3.89	4.02	4.01	0.40	0.36	0.53	0.38	6.89
MuLaCover	8.31	6.27	7.66	7.91	4.54	4.54	4.29	4.39	4.43	0.12	0.42	0.51	0.55	3.36

Appendix

Table 8: Partial-reference cover, outpainting configuration. Same metrics as Table 1 .

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
SongEcho [ 11 ]	7.83	5.90	7.20	7.56	3.50	3.37	3.17	3.27	3.24	0.38	0.35	0.46	0.73	2.62
ACE-Step-1.5 [ 12 ]	8.30	6.16	7.70	7.87	4.31	4.30	4.06	4.16	4.19	0.32	0.32	0.46	0.82	1.68
MuLaCover	8.34	6.33	7.70	7.93	4.58	4.58	4.33	4.43	4.46	0.13	0.41	0.49	0.93	1.28

Appendix

Table 9: Partial-reference cover, inpainting configuration. Same metrics as Table 1 .

Panel of Figure 2 — Figure 2: Listening-study questionnaire interface. Top: welcome page. Middle: demographics page. Bottom: per-case rating page, listing the four metrics together with the textual anchors for scores 1–5.

Background	$n$	Identity Preserveness	Style Controllability	Creativeness	Overall Musicality
Amateur	$200$	$3.54\,[3.27,\,3.79]$	$3.38\,[3.12,\,3.62]$	$3.21\,[3.00,\,3.40]$	$3.50\,[3.35,\,3.67]$
Intermediate	$100$	$3.25\,[2.54,\,3.80]$	$3.08\,[2.37,\,3.70]$	$3.16\,[2.69,\,3.58]$	$3.26\,[2.82,\,3.64]$
Professional	$~80$	$3.26\,[2.82,\,3.88]$	$3.36\,[2.67,\,4.05]$	$3.17\,[2.67,\,3.51]$	$2.89\,[2.28,\,3.34]$

Appendix

Table 10: Mean rating per metric by self-reported musical background, pooled across systems, with cluster-bootstrap

95\%

CI in brackets (resampling at the participant level,

1{,}000

replicates).

n

is the number of (participant, case, system) ratings contributing to each row, per metric.

Metric	$n_{\text{blocks}}$	$\chi^{2}$	$p$ -value	Kendall’s $W$
Identity Preserveness	$95$	$57.19$	$2.3{\times}10^{-12}$	$0.20$
Style Controllability	$95$	$55.69$	$4.9{\times}10^{-12}$	$0.20$
Creativeness	$95$	$52.90$	$1.9{\times}10^{-11}$	$0.19$
Overall Musicality	$95$	$88.08$	$5.7{\times}10^{-19}$	$0.31$

Appendix

Table 11: Friedman omnibus test across the four systems, per metric.

n_{\text{blocks}}

is the number of (participant, case) blocks within which the

k{=}4

system ratings are ranked. Kendall’s

W

is the coefficient-of-concordance effect size.

Pair	Identity	Style	Creativeness	Overall
ACE-Step-1.5 vs. MuLaCover	$1.0000$	$\mathbf{<10^{-4}}$	$\mathbf{0.0004}$	$\mathbf{0.0161}$
ACE-Step-1.5 vs. SongEcho	$\mathbf{<10^{-4}}$	$1.0000$	$\mathbf{0.0169}$	$\mathbf{<10^{-4}}$
ACE-Step-1.5 vs. Suno-v5.5	$\mathbf{0.0475}$	$\mathbf{<10^{-4}}$	$\mathbf{0.0007}$	$\mathbf{<10^{-4}}$
MuLaCover vs. SongEcho	$\mathbf{0.0001}$	$\mathbf{<10^{-4}}$	$\mathbf{<10^{-4}}$	$\mathbf{<10^{-4}}$
MuLaCover vs. Suno-v5.5	$\mathbf{0.0060}$	$1.0000$	$1.0000$	$1.0000$
SongEcho vs. Suno-v5.5	$\mathbf{<10^{-4}}$	$\mathbf{<10^{-4}}$	$\mathbf{<10^{-4}}$	$\mathbf{<10^{-4}}$

Appendix

Table 12: Bonferroni-corrected

p

-values from pairwise Wilcoxon signed-rank tests across the four systems, per metric. Each test pairs the two systems on the within-block (participant, case) score differences over the

n_{\text{blocks}}{=}95

complete blocks; the Bonferroni correction is applied over the

6

pairs within each metric. Bold marks

p<0.05

Attribute	Expert definition / judging criterion
Rhythm and groove	The most salient property of rhythm-and-blues: drums and bass jointly form the groove, and faithful R&B should exhibit typical rhythmic patterns rather than a generic pop or electronic feel.
Vocal melody, repetition, and variation	In R&B, melodic development often happens through modular repetition and varied repetition over a cyclic rhythmic/harmonic frame; phrase-end ornaments, melismas, and vocal timbre are important cues.

Appendix

Table 13: Primary R&B attributes elicited from the expert before any generated audio was played.

Version / system	Rhythm and groove	Vocal melody / timbre	Overall comment
A / ACE-Step-1.5	$2$ : Electric drums, 808, and synthetic bass make the version sound closer to electronic pop.	$2$ : The vocal line is closer to a normal pop-song melody than to R&B melodic development.	The first two attributes deviate substantially from typical R&B; the version leans toward electronic pop rather than canonical R&B.
B / SongEcho	$3$ : The drum timbre and groove contain pop-R&B characteristics.	$3$ : The vocal development is closer to R&B: it uses a four-phrase loop rather than the linear emotional buildup typical of pop ballads.	Overall, this version is relatively better than Version A.
C / MuLaCover	$4$ : The expert rated the rhythm and groove higher than Versions A and B.	$4$ : The expert rated the vocal melodic treatment higher than Versions A and B; the verse also has a vocal timbre that sounds very R&B-like.	The expert judged it close to Version B in broad style but with a more R&B-like singer timbre; it is still not the most canonical R&B rendition.
D / Suno-v5.5	$4.5$ : The groove is slightly better than the preceding versions.	$4.5$ : The vocal line is much more natural and close to pop R&B, especially in phrase endings, melismas, and timbre.	The expert also noted that the auxiliary harmonic and instrumentation-related attributes were relatively well matched.

Appendix

Table 14: Organized expert record for the R&B case study. Each attribute follows the pre-listening criteria in Table 13 . The table focuses on the two primary attributes for which the expert gave complete scores.

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style		CSI judge
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
SongEcho [ 11 ]	7.66	6.26	7.06	7.32	3.47	3.23	3.13	3.19	3.27	0.42	0.32	0.43	0.97	1.03
ACE-Step-1.5 [ 12 ]	8.27	6.08	7.67	7.86	4.37	4.36	4.15	4.25	4.25	0.25	0.31	0.31	0.98	1.08
MuLaCover	8.34	6.24	7.73	7.92	4.52	4.51	4.27	4.37	4.40	0.12	0.40	0.45	0.98	1.02

Appendix

Table 15: Objective evaluation on the full song cover task (Task 1) on human-recordings.

\uparrow

\downarrow

indicate higher / lower is better. Bold marks the best.

	AudioBox [ 54 ] ( $\uparrow$ )				SongEval [ 55 ] ( $\uparrow$ )					Lyrics	Style
Method	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )
HeartMuLa [ 10 ]	8.23	6.32	7.61	7.89	4.68	4.69	4.51	4.58	4.60	0.15	0.27	0.59
MuLaCover -nc	8.30	6.18	7.68	7.92	4.60	4.59	4.34	4.43	4.47	0.11	0.41	0.59

Appendix

Table 16: Comparison between MuLaCover -nc and HeartMuLa [ 10 ] .

\uparrow

\downarrow

indicate higher / lower is better. Bold marks the better result in each column.

Corruption		AudioBox ( $\uparrow$ )				SongEval ( $\uparrow$ )					Lyrics	Style		CSI judge
$p_{1}$	$p_{2}$	PQ	PC	CE	CU	Coh.	Mem.	Nat.	Cla.	Mus.	PER ( $\downarrow$ )	MuQ-Sim ( $\uparrow$ )	MuQ-Dist ( $\uparrow$ )	CSI-R@1 ( $\uparrow$ )	CSI-Rank ( $\downarrow$ )
0	0	8.34	6.34	7.70	7.93	4.59	4.59	4.34	4.44	4.47	0.13	0.41	0.48	0.97	1.05
0	5%	8.29	6.26	7.67	7.90	4.54	4.52	4.25	4.35	4.40	0.13	0.40	0.56	0.63	2.85
0	10%	8.26	6.22	7.61	7.87	4.48	4.46	4.18	4.29	4.34	0.14	0.40	0.57	0.42	5.10
5%	0	8.26	6.20	7.61	7.89	4.51	4.50	4.23	4.32	4.37	0.13	0.39	0.56	0.62	2.91
10%	0	8.29	6.26	7.70	7.91	4.57	4.56	4.28	4.39	4.44	0.11	0.38	0.57	0.44	3.69

Appendix

Table 17: Robustness of MuLaCover to lead-sheet extraction errors on Task 1.

p_{1}

and

p_{2}

denote the probabilities of note removal and pitch shifting, respectively.

\uparrow

\downarrow

indicate higher / lower is better.

	Lead sheet
Method	MS ( $\uparrow$ )	CS ( $\uparrow$ )
SongEcho [ 11 ]	0.77	0.74
ACE-Step-1.5 [ 12 ]	0.63	0.86
Suno-v5.5 [ 53 ]	0.57	0.82
MuLaCover	0.68	0.89

Appendix

Table 18: Lead sheet metrics on the full-song cover task (Task 1).

\uparrow

indicates higher is better. Bold marks the best and underline the second-best result in each column.

Music Generation

Music generation studies computational methods for composing, synthesizing, and editing music from symbolic, audio, textual, and multimodal inputs.

See all 107 papers

Feb 23, 2026

SongEcho: Towards Cover Song Generation via Instance-Adaptive Element-wise Linear Modulation

Cover songs constitute a vital aspect of musical culture, preserving the core melody of an original composition while reinterpreting it to infuse novel emotional depth and thematic emphasis. Although prior research has explored the reinterpretation of instrumental music through melody-conditioned text-to-music models, the task of cover song generation remains largely unaddressed. In this work, we reformulate our cover song generation as a conditional generation, which simultaneously generates new vocals and accompaniment conditioned on the original vocal melody and text prompts. To this end, we present SongEcho, which leverages Instance-Adaptive Element-wise Linear Modulation (IA-EiLM), a framework that incorporates controllable generation by improving both conditioning injection mechanism and conditional representation. To enhance the conditioning injection mechanism, we extend Feature-wise Linear Modulation (FiLM) to an Element-wise Linear Modulation (EiLM), to facilitate precise temporal alignment in melody control. For conditional representations, we propose Instance-Adaptive Condition Refinement (IACR), which refines conditioning features by interacting with the hidden states of the generative model, yielding instance-adaptive conditioning. Additionally, to address the scarcity of large-scale, open-source full-song datasets, we construct Suno70k, a high-quality AI song dataset enriched with comprehensive annotations. Experimental results across multiple datasets demonstrate that our approach generates superior cover songs compared to existing methods, while requiring fewer than 30% of the trainable parameters. The code, dataset, and demos are available at https://github.com/lsfhuihuiff/SongEcho_ICLR2026.

Sifei Li, Yang Li, Zizhou Wang +5

MAIS, Institute of Automation, Chinese Academy of Sciences · School of Artificial Intelligence, University of Chinese Academy of Sciences · ISRC, Institute of Software, Chinese Academy of Sciences +2

Music Generation Controllable Music Generation Text-to-Music Generation

Ask PDF

Jul 29, 2026

MPEcho: A Melody and Phoneme-Aware Generative Framework for Controllable Cover Song Generation

Cover song generation (CSG) should preserve the melodic and linguistic content of a reference song while recreating the remaining musical components. The state-of-the-art model SongEcho utilizes

F_0

sequences and voiced/unvoiced (V/UV) tags for conditioning; however, implicit linguistic information from V/UV tags cannot guarantee lyric accuracy, leading to a high phoneme error rate (PER). Inspired by singing voice synthesis (SVS), we propose MPEcho, which integrates a phoneme encoder and a length regulator (LR) into the SongEcho framework. By providing explicit phoneme-level conditioning and precise temporal boundaries, MPEcho significantly reduces PER. To enable this, we developed Phonsa, a Whisper-based automatic transcription model that provides high-precision phoneme-level annotations for singing voices, overcoming the scarcity of high-quality audio-phoneme pairs. Experimental results validate the effectiveness of Phonsa for alignment and MPEcho for end-to-end CSG. The audio samples, code and weights can be accessed from https://lonian6.github.io/MPEcho.github.io/.

Wei-Jaw Lee, Hsuan-Yu Yeh, Ting-Yi Hu +3

GICE, National Taiwan University

Controllable Music Generation Music Generation

Ask PDF

Jul 13, 2026

Qwen-Music Technical Report

We introduce Qwen-Music, a music generation model that produces high-fidelity songs with complete vocals. It supports text-to-music generation from descriptions, lyrics, and musical attributes, and cover song generation with different styles and vocal characteristics. Qwen-Music comprises three components: Qwen-Music-Tokenizer, Qwen-Music-LLM, and Qwen-Music-Render. The tokenizer compresses audio into a 25 Hz single-codebook stream of Music Semantic Tokens that preserve semantic and melodic information. The LLM performs autoregressive modeling with a melody-token-based chain-of-thought (Melody-CoT) mechanism that plans melodies before full-song generation, improving musicality, structural coherence, and reference-melody preservation. The renderer enriches discrete semantic tokens with acoustic details to produce high-fidelity stereo waveforms. We train the LLM using a quality-aware pre-training curriculum followed by progressive post-training with supervised initialization, offline DPO, and online GSPO to improve musicality and instruction following. On 600 evaluation inputs, 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 generation, Qwen-Music preserves reference melodies more accurately than Suno V5.5, Suno V5, and MiniMax Cover on the AI-generated reference set, and outperforms MiniMax Cover on most metrics on the real-world popular-song reference set.

Jin Xu, Kangdi Wang, Ruibin Yuan +24

Qwen Team

Text-to-Music Generation Controllable Music Generation Music Generation

Ask PDF

Abstract

Music Generation

Explore similar work

SongEcho: Towards Cover Song Generation via Instance-Adaptive Element-wise Linear Modulation

MPEcho: A Melody and Phoneme-Aware Generative Framework for Controllable Cover Song Generation

Qwen-Music Technical Report