Human-like Speech Synthesis

The “perfection” of traditional TTS is the problem: overly smooth speech gives away the machine. Human “imperfections” (pauses, “um’s,” repetitions) are an external manifestation of thought. The solution: LLM outputs control tags ([THINKING] – pause with filler, [EMO:happy], [SPEED:0.8x]), TTS acts as a multimodal generator: plain text → speech, tag → corresponding non-speech sound. Only the LLM knows where a pause for deliberation is needed. Implementation: a custom model with native tag support or voice cloning with a library of reference recordings for different emotions and speeds.

Related: [Voice paradigms], [Cascaded voice pipeline]