just general research work. Once the recipes are efficient enough the modality is a smaller detail.
On the product side we're trying to orient more towards 'productive work assistant' rather than the default pull of audio models towards being an 'ai friend'.