Premature audio cutoff

[sproft]

For some reason the audio output of the generated response is always cut short>

For example here the word "away" is not heard.

[SmolinskiP]

Please tell me TTS engine you are using. There was similar issue with some models on Piper TTS

[sproft]

i am running Piper TTS from within HomeOS

Piper
Current version: 2.2.2 (Changelog)

[SmolinskiP]

i am running Piper TTS from within HomeOS

Piper Current version: 2.2.2 (Changelog)

Can you try this version of piper and set the option in last line?

[RheaAyase]

[RheaAyase]

The issue is in my opinion client side = both input and output.

Input: You can start telling it what light to turn on, say half a sentence, then trigger the Assist, and that's the only way it works correctly. It has like 1 second delay basically so you have to call assist mid sentence for it to be fully picked up.

Context: Kokoro TTS and onnx-asr STT in VRAM (Tesla HBM2 you can't any better latency.)

[RheaAyase]

These settings "helped" with the input problem (I have a different output problem so can't test that) - the problem is still present though, even though it's less of an issue.

[SmolinskiP]

The issue is in my opinion client side = both input and output.

Input: You can start telling it what light to turn on, say half a sentence, then trigger the Assist, and that's the only way it works correctly. It has like 1 second delay basically so you have to call assist mid sentence for it to be fully picked up.

Context: Kokoro TTS and onnx-asr STT in VRAM (Tesla HBM2 you can't any better latency.)

I don't think i fully understand. You mean for example something like that:

Turn on light
Alexa
In kitchen?

Or what?

[RheaAyase]

That the input audio is delayed. I can say a sentence, THEN trigger the app, and it still picked up that sentence i said BEFORE triggering it. Which is also why it's "prematurely cut off" because you're done talking, it stops receiving, but because it's delayed by a second or two it cuts off whatever "didn't arrive yet due to the delay"

[SmolinskiP]

@RheaAyase I added flushing audio buffer after wake word activation - on windows (WASAPI ) it was about 50ms, so not visible but on Linux (ALSA/PulseAudio) it can make difference. Flatpak image is building, please check the fix when it finishes (or just run from source)