Sentence: Rattle is the sound a [MASK] makes.
Answer: Snake
BERT: Dog
ITH: Snake
Sentence: The sound of chirping in the fall is often associated with a [MASK].
Answer: Cricket
BERT: Bird
ITH: Cricket
Sentence: The sound of guitar is [MASK] than the sound of sopranissimo saxophone.
Answer: Lower
w/o DKI: Higher
ITH: Lower
Sentence: The sound of alto trombone is [MASK] than the sound of sopranino trombone.
Answer: Lower
w/o DKI: Higher
ITH: Lower
Language models pretrained on text-only corpora often struggle with tasks that require auditory commonsense knowledge. Previous work addresses this problem by augmenting the language model to retrieve knowledge from external audio databases. This approach has several limitations, such as the potential lack of relevant audio in databases and the high costs associated with constructing and querying the databases. To address these issues, we propose Imagine to Hear, a novel approach that dynamically generates auditory knowledge using generative models. Our framework detects multiple audio-related textual spans from the given prompt and generates corresponding auditory knowledge. We develop several mechanisms to efficiently process multiple auditory knowledge, including a CLAP-based rejection sampler and a language-audio fusion module. Our experiments show that our method achieves state-of-the-art performance on AuditoryBench without relying on external databases, highlighting the effectiveness of our generation-based approach.
An illustration of the overall framework of the proposed Imagine to Hear (ITH), consisting of three components:
1) An imagination module, which detects multiple audio-related spans from the given prompt and generates multiple corresponding audio knowledge.
2) A fusion module, which combines the (variable-length) auditory and textual information.
3) A language encoder, which processes the output of the fusion module.
TBD