3 comments

[ 2.7 ms ] story [ 21.1 ms ] thread
Hello!

I ported Suno AI's Bark text-to-speech model in C/C++ to allow fast, realistic, multilingual audio generation on the CPU.

Generating a 5-second audio with vanilla Bark takes 1 minute on a M1 Pro CPU. Using my port in C++ with ggml, it goes down to 15 seconds.

I aim to bring it down to a second to allow on-device real-time audio generation.