I have a robotic and virtual agent project that I'm trying to get as close to real-time response as possible.
I use the following to generate speech:
python3 fastVoice.py | larynx -v ek --interactive --ssml --raw-stream --cuda --half --max-thread-workers 8 --stdin-format lines --process-on-blank-line| aplay -r 22050 -c 1 -f S16_LE
Where fastVoice.py just dumps the SSML from a socket onto stdin (remember to flush properly ...)
fastVoice.txt
All works very well. Audio generally starts <1s from receiving the message. The question is how to get a phoneme-viseme sequence synced with the audio output.
I can manage to get level 0-ish lipsync by looking at the amplitude of the audio output, but that gives enough info for just the jaw, not the viseme's of the lips.
Do you have any ideas/pointers on how to maintain the responsiveness of "--raw-stream" while getting real-time matching info to generate the matching visemes?
First, love the project !
I have a robotic and virtual agent project that I'm trying to get as close to real-time response as possible. I use the following to generate speech:
python3 fastVoice.py | larynx -v ek --interactive --ssml --raw-stream --cuda --half --max-thread-workers 8 --stdin-format lines --process-on-blank-line| aplay -r 22050 -c 1 -f S16_LEWhere fastVoice.py just dumps the SSML from a socket onto stdin (remember to flush properly ...) fastVoice.txtAll works very well. Audio generally starts <1s from receiving the message. The question is how to get a phoneme-viseme sequence synced with the audio output. I can manage to get level 0-ish lipsync by looking at the amplitude of the audio output, but that gives enough info for just the jaw, not the viseme's of the lips.
Do you have any ideas/pointers on how to maintain the responsiveness of "--raw-stream" while getting real-time matching info to generate the matching visemes?