Soul-AILab/SoulX-Singer-Eval-Dataset

SoulX-Singer-Eval

This corpus contains 100 singing segments from 50 distinct individuals (25 Mandarin and 25 English speakers), with 2 segments provided per speaker. Additionally, 30 target samples are cross-selected from the Opencpop, M4Singer, and GTSinger, which are also in GMO-SVS. The annotation files are organized in JSONL format and categorized by text formulation (phoneme or word) and usage (prompt or target). All waveforms are stored in the audio directory. Metadata includes source dataset, song title, singer ID, language, and audio paths.

Word-based annotations include note_text, note_dur, note_pitch, note_type, all labeled at the individual singing note level. note_text specifies the lyric word, while note_dur, note_pitch and note_type denote the duration, pitch class, and category (1 for rest, 2 for lyric, and 3 for slur) of each note. Phone-based annotations follow the GTSinger style, including ph, ep_pitches, ep_notedurs, ep_types and ph_durs.

GMO-SVS

This corpus chooses 802 samples from Opencpop, M4Singer, and GTSinger. The audio and annotation files follow the same organizational structure as the SoulX-Singer-Eval. Metadata, word-based annotations, and phone-based annotations are all strictly aligned with the SoulX-Singer-Evalset specifications to ensure consistency.