Autoregressive Speech Synthesis without Vector Quantization