We learn matrix representations for the fre- quent sound-relevant adjectives of English and compose them with vector representations of their nouns. The matrices are learnt jointly from audio and textual data, via linear regres- sion and tensor skipgram. They are assessed using an adjective similarity benchmark and also a novel adjective-noun phrase similarity dataset, applied to two tasks: semantic similar- ity and audio similarity. Joint learning via Ten- sor Skipgram (TSG) outperforms audio-only models, matrix composition outperforms addi- tion and non compositional phrase vectors.