Papers
Topics
Authors
Recent
Search
2000 character limit reached

MADVSD: Multi-Accent Mandarin Vocal Dataset

Updated 14 December 2025
  • MADVSD is a large-scale dataset featuring 670 hours of unprocessed vocal recordings from 4,206 Mandarin speakers across nine distinct Chinese regions.
  • It combines standardized phonetic exercises and self-selected pop song recordings, processed with noise reduction and filtering to facilitate fine-grained accent analysis.
  • Benchmark results reveal its effectiveness for singing accent recognition and cross-domain accent modeling, supporting AI-driven music and sociolinguistic studies.

The Multi-Accent Mandarin Dry-Vocal Singing Dataset (MADVSD) is a large-scale corpus specifically designed to facilitate research in singing accent recognition—the study of regional accent manifestation in sung Mandarin, which remains underexplored due to the lack of suitable data. MADVSD addresses critical limitations of prior singing datasets, providing over 670 hours of “dry” (a cappella, unprocessed) vocal data from 4,206 native Mandarin speakers across nine distinct Chinese regions. Each participant contributed three popular song recordings and a standardized phonetic exercise designed for fine-grained analysis of pitch and vowel variation. This dataset enables benchmarking of speech models in singing contexts and supports the examination of dialectal influences and phonetic features relevant to accentual variation in music (Wang et al., 7 Dec 2025).

1. Dataset Structure and Regional Coverage

MADVSD consists of two principal recording types from each speaker: three popular Mandarin song performances (self-selected from genres including ballad pop, rock-infused pop, and folk ballads; totaling roughly 8–12 minutes per speaker), and a phonetic exercise session (covering all Mandarin vowel phonemes as sustained notes and a full octave scale: Do, Re, Mi, Fa, So, La, Ti, Do). The cumulative dataset incorporates 12,618 song clips plus 4,206 vowel exercise clips, amounting to 670 hours of dry vocal data.

Speakers are stratified into nine geographic regions for accent annotation, balancing sample size and broad coverage. The regional breakdown is detailed below:

Region (Abbrev.) Speaker Count
Shanghai–Zhejiang Region (SZR) 782
Jiangsu Region (JR) 290
Shanxi–Shandong–Henan Region (SSHR) 362
Beijing–Inner Mongolia–Liaoning Region (BMLR) 614
Shaanxi–Gansu–Xinjiang Region (SGXR) 400
Fujian Region (FR) 455
Hunan–Hubei–Anhui Region (HHAR) 438
Guangdong–Guangxi Region (GGR) 484
Yunnan–Guizhou–Sichuan Region (YGSR) 381

MADVSD does not provide fine-grained sub-dialect annotations; only broad region labels are supplied. Metadata includes speaker ID, region, province, city, and unconstrained device type (typically mobile phones). Age and gender were not systematically recorded.

2. Data Collection and Preprocessing Protocol

Participants recorded sessions in quiet home environments, often utilizing makeshift vocal booths (pillows, quilts) to control acoustics and avoid noise. Recording protocol required headphone-monitored instrumentals for synchronization while preserving purely a cappella vocals.

Audio preprocessing comprised:

  1. Stereo separation and channel selection based on amplitude, followed by conversion to mono.
  2. Noise reduction and dereverberation using LALAL.AI and UVR5.
  3. High-pass filtering at 80 Hz to remove subsonic rumble.
  4. Final output as single-channel, 16-bit WAV files.

Annotation employed both ASR and LLM-based parsing to extract accent region, province, and city from speakers’ declarative hometown utterance at session start. Quality control procedures included guideline enforcement for environment and recording checks against clipping and plosives.

3. Accent Recognition Benchmarks

MADVSD supports benchmarking for both 9-way Mandarin region accent classification (MR-ACC) and binary region accent detection (AD). Models evaluated include generic audio encoders (MelGAN, VGGish, YAMNet, Wav2Vec2.0, HuBERT) and speech accent specialists (KeSpeech, TeleSpeech-Pretrain-L, Qifusion, DIMNet).

Metrics are formally specified as: Accuracy=TP+TNTP+TN+FP+FN,Precision=TPTP+FP,Recall=TPTP+FN,F1=2  P×RP+R\text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN},\quad \text{Precision} = \frac{TP}{TP + FP},\quad \text{Recall} = \frac{TP}{TP + FN},\quad F1 = 2\;\frac{P \times R}{P + R} Primary reporting emphasizes classification accuracy.

Empirical results demonstrate that DIMNet achieves the highest overall multi-region accuracy and dominates most binary detection tasks, with Qifusion providing optimal performance on BMLR-AD and YGSR-AD. Binary AD accuracies frequently exceed 80%, while MR-ACC approximates 60%, reflecting greater fine-grained discrimination difficulty. Accent features in regions such as GGR, YGSR, SZR, and SGXR are particularly distinctive, with specialist models (KeSpeech, DIMNet, Qifusion) exceeding generic pre-trained encoders. All models, however, show reduced performance in singing versus speech contexts, signifying the gap between spoken and sung accent modeling.

4. Dialectal Influence and Phonetic Analysis

Dialectal influence in singing accent recognition was quantified by training KeSpeech’s sub-dialect model on distinct regimes:

Training Data MR-ACC
KeSpeech only 51.3 %
MADVSD only 56.8 %
KeSpeech + MADVSD 60.1 %

Joint training yields superior MR-ACC, confirming the complementarity of spoken dialect data and substantiating the dialect-to-accent influence hypothesis.

Phonetic analysis leveraged DIMNet embeddings and Montreal Forced Aligner to define the Vowel Analysis Metric (VAM): cosine similarity between vowel embeddings from different regions, where lower values indicate greater phonetic divergence. Comparative VAM scores for YGSR versus SSHR vowels revealed pronounced accentual differences in /u/ (0.45), /aʊ/ (0.48), and /aɪ/ (0.52), with minimal divergence in /o/ (0.78), /i/ (0.75), and /y/ (0.71). Diphthongs and the high back rounded vowel /u/ emerged as key regional accent markers in Mandarin singing.

5. Access, Usage, and Research Integration

MADVSD is available for non-commercial research by request at https://github.com/CarlWangChina/MADVSD, subject to application approval and data-use agreement. Audio and metadata are then released.

Potential applications of MADVSD include:

  • Singing accent recognition and inference of geographic origin
  • Singing accent conversion (cross-accent voice transformation)
  • Accent-controlled singing synthesis
  • Phonetic and sociolinguistic musical performance studies

Limitations encompass genre restriction to contemporary Mandarin popular styles, regional annotation at the broad (nine-category) level, absence of speaker age/gender metadata, and heterogeneous recording device quality. Expansion to rap, opera, folk, and classical genres, finer accent categorization, and increased metadata granularity represent future improvement pathways.

6. Significance and Future Directions

MADVSD constitutes the largest publicly documented dry-vocal singing dataset with annotated Mandarin regional accents. It enables rigorous research in singing accent recognition, cross-domain accent modeling, and AI-music applications. Foreseeable extensions include singing accent conversion modeling, perceptual studies on accent aesthetics, and progressively granular accent mapping to provinces or cities. Ongoing research will likely further delineate the relationships between spoken dialects and vocal musical characteristics, underpinning advances in both computational and musicological studies (Wang et al., 7 Dec 2025).

Definition Search Book Streamline Icon: https://streamlinehq.com
References (1)

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to Multi-Accent Mandarin Dry-Vocal Singing Dataset (MADVSD).