RTF && FFT?

[HaiFengZeng]

cosyvoice2:by deploying with vllm, the first chunk (15 tokens) of LLM model will take about 75ms and with flow + hift model, it can get a first frame time about 150ms, so with rwkv, with same 15 tokens, how long will it take and how many resoures will use?

[yynil]

I didn't implement the streaming mode. However with the newest rwkv-fla package in 4090, RTF is around 0.5.
For very long sentences generation, RWKV TTS can continue to generate sentence by sentence in a "fake" streaming mode. Please refer the long generation example: https://github.com/yynil/RWKVTTS/blob/main/model/test/test_instructed.py by passing a long text file as an argument.