Organizations: School of Information Management, Wuhan University, Wuhan, China · Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, Wuhan, China · ShenYang Conservatory Of Music, ShenYang, China. · School of Computer Science, Wuhan University, Wuhan, China
Recently, there have been significant advancements in music generation. However, existing models primarily focus on creating modern pop songs, making it challenging to produce ancient music with distinct rhythms and styles, such as ancient Chinese SongCi. In this paper, we introduce SongSong, the first music generation model capable of restoring Chinese SongCi to our knowledge. Our model first predicts the melody from the input SongCi, then separately generates the singing voice and accompaniment based on that melody, and finally combines all elements to create the final piece of music. Additionally, to address the lack of ancient music datasets, we create OpenSongSong, a comprehensive dataset of ancient Chinese SongCi music, featuring 29.9 hours of compositions by various renowned SongCi music masters. To assess SongSong's proficiency in performing SongCi, we randomly select 85 SongCi sentences that were not part of the training set for evaluation against SongSong and music generation platforms such as Suno and SkyMusic. The subjective and objective outcomes indicate that our proposed model achieves leading performance in generating high-quality SongCi music.
Figures & tables
Figure 1: A piece of ancient Chinese SongCi music recorded using Chinese Gongche notation, which is vertically arranged and uses special Chinese symbols as musical notes. The blue box indicates the SongCi, and the red box indicates the notes.
Figure 2: The relationship between lyric, rhythm, and melody.
Figure 3: The structure of our proposed model, SongSong. The English meaning of input SongCi is “How rare the moon,so round and clear! With cup in hand,I ask of the blue sky.”
Figure 4: The design of the singing voice generation module.
Figure 5: The specific architecture of adopted accompany generation module.
Dataset
Style
Hours
Singers
Phoneme
Pitch
MIDI
Annotation
Text
Phoneme
Pitch
Alignment
Opencpop
pop
5.25
1
miss “van”,“ve”,“vn”
D#3-D#5
✓
✓
✓
✓
✓
M4Singer
pop
29.77
20
no miss
G1-G5
✓
✓
✓
✕
✓
OpenSongSong
SongCi
29.90
89
no miss
G1-C6
✓
✓
✓
✓
✓
Table 1: The comparison results for OpenSongSong, Opencpop, and M4Singer. The “Phenome” indicates if certain phonemes are absent from the dataset, while the “Pitch” reflects the distribution of pitches within the dataset. Phonemes or pitches that occur fewer than 20 times are deemed excluded from the dataset. The “MIDI” indicates whether the dataset includes music score information. The “Annotation” specifies the number of annotations present in the dataset. Lastly, the “Alignment” indicates whether each annotation is synchronized with the corresponding singing voice recording, allowing for the start and end times of each element to be extracted from the annotation type.
Test
Model
Objective Metrics
Common Subjective Metrics
SongCi Subjective Metrics
FADvgg ( ↓ )
FADpann ( ↓ )
MS ( ↑ )
ER ( ↑ )
MC ( ↑ )
SQ ( ↑ )
PA ( ↑ )
VN ( ↑ )
SCS ( ↑ )
AC ( ↑ )
Zero-shot
Suno
5.74
3.35e-4
65.48
68.85
63.19
62.96
62.59
65.37
35.30
19.28
SkyMusic
7.88
3.42e-4
63.96
65.48
59.11
72.67
54.81
72.89
25.37
45.69
SongSong
5.41
3.87e-4
56.67
46.26
53.93
55.56
75.37
65.56
78.44
65.54
Few-shot
Suno
7.92
3.98e-4
64.44
62.30
64.41
75.15
48.85
71.04
42.19
28.90
SkyMusic
6.31
5.55e-5
60.26
58.30
60.63
76.41
59.04
72.74
46.67
52.60
Table 2: The comparison results for SongSong, Suno, and SkyMusic. The terms zero-shot and few-shot indicate whether Songci audio was provided as a reference for Suno and SkyMusic. For subjective metrics, MS stands for music structure, ER denotes equipment richness, MC indicates motivation continuity, PA refers to pronunciation accuracy, VN signifies voice naturality, SQ represents sound quality, SCS pertains to conformity with the SongCi style and AC is acompaniment conformity.
Figure 6: The mel-spectrogram comparisons among ground truth, SongSong, Suno and Skymusic.
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.
Recent advancements in Text-to-Song generation have enabled realistic musical content production, yet existing evaluation benchmarks lack the professional granularity to capture multi-dimensional aesthetic nuances. In this paper, we propose SongBench, a specialized framework for fine-grained song assessment across seven key dimensions: Vocal, Instrument, Melody, Structure, Arrangement, Mixing, and Musicality. Utilizing this framework, we construct an expert-annotated database comprising 11,717 samples from state-of-the-art models, labeled by music professionals. Extensive experimental results demonstrate that SongBench achieves high correlation with expert ratings. By revealing fine-grained performance gaps in current state-of-the-art models, SongBench serves as a diagnostic benchmark to steer the development toward more professional and musically coherent song generation.
Dapeng Wu, Shun Lei, Wei Tan +5
Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Tencent
Persian music, with its unique tonalities, modal systems (Dastgah), and rhythmic structures, presents significant challenges for music generation models trained primarily on Western music. We address this gap by curating the first large-scale dataset of Persian songs, comprising over 900 hours high-quality audio samples across diverse sub-genres, including pop, traditional, and contemporary styles. This dataset captures the rich melodic and cultural diversity of Persian music and serves as the foundation for fine-tuning MusicGen, a state-of-the-art generative music model. We adapt MusicGen to this domain and evaluate its performance by utilizing subjective and objective metrics. To assess the semantic alignment between generated music and intended style tags, we report the proportion of relevant tags accurately reflected in the generated outputs. Our results demonstrate that the fine-tuned model produces compositions that more align with Persian stylistic conventions. This work introduces a new resource for generative music research and illustrates the adaptability of music generation models to underrepresented cultural and linguistic contexts.
Mohammad Hossein Sameti, Diba Hadi Esfangereh, Sepehr Harfi Moridani +2
Sharif University of Technology · Independent Researcher