Currently, the TextToSpeech pipeline generates output speech as a single batch. In order to support real-time use cases, similar to streaming LLM generation, this pipeline should be modified to yield chunks of audio.