Основні поняття
Seed-Music is a versatile framework that leverages both auto-regressive language modeling and diffusion approaches to enable high-quality music generation with fine-grained style control, as well as interactive editing of generated music.
Анотація
Seed-Music is a comprehensive framework for music generation and editing that addresses the challenges in this domain:
-
Domain Complexity: Seed-Music's unified approach handles the complexity of music signals, including melodic coherence, structural consistency, overlapping sounds, and expressive singing techniques.
-
Evaluation Difficulty: Seed-Music employs a combination of quantitative metrics (e.g., WER, MIR classification) and qualitative assessments (e.g., Comparative Mean Opinion Score) to evaluate the musicality, audio quality, and prompt adherence of the generated outputs.
-
Data Complexity: Seed-Music utilizes various intermediate representations, including audio tokens, symbolic music tokens, and vocoder latents, to effectively leverage different types of annotated music data.
-
Diverse User Needs: Seed-Music supports multiple workflows, from controlled music generation to post-production editing, catering to the needs of both novice and professional users.
The framework consists of three core components: a Representation Learning module, a Generator, and a Renderer. The choice of intermediate representation (audio tokens, symbolic tokens, or vocoder latents) depends on the specific downstream task.
Key applications of Seed-Music include:
- Lyrics2Song: Generating high-quality vocal music from lyrics and style descriptions.
- Lyrics2Leadsheet2Song: Generating music from lyrics, with an interpretable symbolic representation (lead sheet tokens) as an intermediate step.
- MusicEDiT: Enabling interactive editing of lyrics and melodies in existing music audio.
- Zero-shot Singing Voice Conversion: Allowing users to modify the timbre of vocals in existing audio using only a short speech recording as reference.
Seed-Music aims to lower the barriers to music creation and empower both novices and professionals to engage in different stages of the music production process.
Статистика
"Music signals are highly complex, exhibiting both short-term melodic coherence and long-term structural consistency."
"Evaluating music generation models often requires domain expertise to assess artistic quality, including judging the appeal of melodies, the coherence of chord progressions, the presence of idiomatic structure, and the expressiveness of vocals."
"Music annotation requires special domain knowledge. While many people can transcribe speech or label images, identifying musical elements such as chords, song sections, instruments, and genres requires a strong musical background."
"The needs of novice musicians differ greatly from those of professionals. A text-to-music system that generates a complete audio piece from a language prompt can be transformative for a beginner, but may offer limited value to professional producers, who typically seek more granular control over compositions and access to individual instrument stems."
Цитати
"Music is deeply embedded in human culture. Throughout human history, vocal music has accompanied key moments in life and society: from love calls to seasonal harvests."
"Our goal is to leverage modern generative modeling technologies, not to replace human creativity, but to lower the barriers to music creation. By offering interactive creation and editing tools, we aim to empower both novices and professionals to engage at different stages of the music production process."