OAI compatible clone/synthesize server

[khimaros]

i'm maintaining here https://github.com/khimaros/s2.cpp

[tyllmoritz]

I was trying to use the playground from llama-swap and it did not detect the voice list.

Steps:
config.yaml for llama-swap

models:
  s2-pro-q4_k_m:
     cmd: /home/user/git/s2.cpp/build/s2-server --vulkan 0 -m /home/user/git/s2.cpp/models/s2-pro-q4_k_m.gguf --host 127.0.0.1 --port ${PORT}

add voices to running s2.cpp with curl POST requests:

curl -X POST http://127.0.0.1:8080/upstream/s2-pro-q4_k_m/v1/audio/voices --form "audio_sample=@/home/user/git/s2.cpp/voices/chinese_woman.wav" --form "ref_text='希望你以后能够做得比我还好哟'" --form "name=chinese_woman"

list available voices:

curl -X GET http://127.0.0.1:8081/upstream/s2-pro-q4_k_m/v1/audio/voices
{"s2-pro-q4_k_m":["default","voice_1"]}

in the llama-swap playground:

• select the model in the 'Speech' tab
• edit the voice list with browser developer tools
• switch back and forth to select voice 'voice_1'
• generate

I'll try to see what is llama-swaps problem with the GET response from /v1/audio/voices

Is it possible to set a custom id with the POST request, so that it does not always use voice_1, voice_2, etc in the response?

Edit: llama-swap wants the response to be "{"voices":["default","voice_1"]}" instead of "{"s2-pro-q4_k_m":["default","voice_1"]}". It works now (I just renamed my gguf file to voices.gguf)