Improve Azure Speech TTFB by reusing connection at each voice synthesis

[yousri-sellami]

The Azure Speech SDK uses a websocket to communicate with the service. If the connection is newly established, the network latency includes extra time to establish the connection.

In my experiments with the Python SDK, TTFB was divided by ~5 when reusing the SpeechSynthesizer as Azure recommends.

I have tried to move the Azure speech synthesizer as a TTS attribute in order to instanciate once and dynamically update its audio config with each stream. Unfortunately, this doesn't reduce latency. As David suggested, it looks like we would need to keep one stream to make sure the connection is reused.