It seems like to actually achieve streaming you'd need to stream not just the output audio bytes but also the text to be vocalized.
E.g. if you have a language model determine what the agent is going to say, and if we want to achieve the lowest latency possible, shouldn't we immediately feed those tokens to csm as they are streaming out of the language model? I.e. the input text for csm-streaming needs to be a generator of some sort rather than a fixed string?
It seems like to actually achieve streaming you'd need to stream not just the output audio bytes but also the text to be vocalized.
E.g. if you have a language model determine what the agent is going to say, and if we want to achieve the lowest latency possible, shouldn't we immediately feed those tokens to csm as they are streaming out of the language model? I.e. the input text for csm-streaming needs to be a generator of some sort rather than a fixed string?