Real-Time Audio-to-Score Alignment of Singing Voice Based on Melody and Lyric Information

Rong Gong (1) , Philippe Cuvillier (1) , Nicolas Obin (2) , Arshia Cont (3, 1)

1 MuTant - Synchronous Realtime Processing and Programming of Music Signals
2 Analyse et synthèse sonores [Paris]
3 Repmus - Représentations musicales

Singing voice is specific in music: a vocal performance conveys both music (melody/pitch) and lyrics (text/phoneme) content. This paper aims at exploiting the advantages of melody and lyric information for real-time audio-to-score alignment of singing voice. First, lyrics are added as a separate observation stream into a template-based hidden semi-Markov model (HSMM), whose observation model is based on the construction of vowel templates. Second, early and late fusion of melody and lyric information are processed during real-time audio-to-score alignment. An experiment conducted with two professional singers (male/female) shows that the performance of a lyrics-based system is comparable to that of melody-based score following systems. Furthermore, late fusion of melody and lyric information substantially improves the alignment performance. Finally, maximum a posteriori adaptation (MAP) of the vowel templates from one singer to the other suggests that lyric information can be efficiently used for any singer.

Mots clés: Index Terms: singing voice; real-time audio-to-score alignment; lyrics; spectral envelope; information fusion; singer adaptation

Domaines: Traitement du signal et de l'image [eess.SP]; Machine Learning [stat.ML]; Son [cs.SD]

Rong Gong, Philippe Cuvillier, Nicolas Obin, Arshia Cont. Real-Time Audio-to-Score Alignment of Singing Voice Based on Melody and Lyric Information. Interspeech, Sep 2015, Dresde, Germany. ⟨hal-01164550⟩