MIDI-Conditional Text-to-Audio Synthesis Using ControlNet on AudioLDM

Text-to-audio systems have gained attention in recent months, achieving impressive results in general audio synthesis. However, they often lack fine-grained control over the musical output, as note-level adjustments cannot be determined by text. In this work, we present MIDI-AudioLDM, which implemen...

Descripción completa

Detalles Bibliográficos
Autor: Ibáñez Martínez, Laura
Tipo de recurso: tesis de maestría
Fecha de publicación:2023
País:España
Institución:Universidad Nacional de Educación a Distancia
Repositorio:e-spacio. Repositorio Institucional de la UNED
Idioma:inglés
OAI Identifier:oai:e-spacio.uned.es:20.500.14468/23782
Acceso en línea:https://hdl.handle.net/20.500.14468/23782
Access Level:acceso abierto
Palabra clave:1203.17 Informática
audio synthesis
MIDI conditioning
text-to-audio systems
AudioLDM
ControlNet
Descripción
Sumario:Text-to-audio systems have gained attention in recent months, achieving impressive results in general audio synthesis. However, they often lack fine-grained control over the musical output, as note-level adjustments cannot be determined by text. In this work, we present MIDI-AudioLDM, which implements MIDI conditioning into AudioLDM with the use of ControlNet. This enables MIDI-conditional text-to-audio synthesis, which adds up to AudioLDM’s previous capacities, including direct text-to-audio synthesis as well as audio style transfer and inpainting. Like AudioLDM, the model uses contrastive language-audio pretraining (CLAP) latents and is trained on audio embeddings, while using text embeddings for inference. In contrast to unconditional audio synthesis, MIDI-AudioLDM offers detailed control over various musical aspects such as notes, genre, mood, and timbre, which makes it a more valuable tool for the music production process. A demo is available at https://huggingface.co/spaces/lauraibnz/midi-audioldm.