Warning, truncation.

[onehundredfeet]

I am getting this while running it:

WARNING:root:frame length (1103) is greater than FFT size (512), frame will be truncated. Increase NFFT to avoid.

My file is 44100 hz, as the original one states.

[fadiaburaid]

I am getting the same warning as well even with the sample files. However, I don't know whether it affects the output. I don't use Maya so I have no way to test it. I was planning to use Three.js to control the rig blendshapes but this will take some time.

[jh-gglabs]

Hi, I have found a related issue. The time at converting to speech feature using python_speech_features would make some problems.

[fadiaburaid]

After testing the viseme output on actual avatar rig, I can confirm that the warning has no effect. Just make sure the input wav file has a sample rate of 44100hz. The results are not the best because I am using an avatar with Oculus viseme format not a JALI rig so I had to do a conversion. I am sure I can improve on that.

https://github.com/junhwanjang/visemenet-inference/assets/28487342/170ce2f0-a415-41c2-90be-a1067bc1c9e5

[jh-gglabs]

@fadiaburaid Thanks for sharing, it looks nice though, can you share how to do the conversion from JALI to Oculus viseme format?

[fadiaburaid]

@jh-gglabs I just map the visemes to their Oculus equivalent ones (Look at the table below). If multiple visemes with the same representation fire up at the same time I take the average. Something still doesn't seem right I am still experimenting. I wanted to use Ovrlipsync as it has better performance from my experience with it in Unity. I managed to find the model, but I am still unable to view the model graph and do the inference.

index	JALI	Oculus Representation
0	Jaw	Not used
1	Lip	Not used
2	Ah	aa
3	Aa	aa
4	Eh	E
5	Ee	E
6	Ih	I
7	Oh	O
8	Uh	U
9	U	U
10	Eu	E
11	Schwa	E
12	R	RR
13	S	SS
14	Sh Ch Zh	CH
15	Th	TH
16	JY	CH
17	LNTD	RR
18	GK	kk
19	MBP	PP
20	FV	FF
21	WA_PEDAL	Not Used

[junhwanjang]

Thanks for sharing :)