# [audio.cpp] SoulX-FlashHead + LFM3-Audio + OWSM v4 = Real-Time Local Avatar Chat — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/19361bb52ed2cd70

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-15 18:29 ingested 2026-11-16 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xx2vi0/audiocpp_soulxflashhead_lfm3audio_owsm_v4/)

## retained text



```
In audio.cpp 0.9.5 you can try this new avatar voice chat "app" in the WebUI.

The pipeline combines 3 models:  SoulX-FlashHead for real-time talking head generation,  OWSM v4 for transcribing speech input to text, and LFM3-Audio for conversational responses and TTS.

LFM3-Audio is actually called twice. The speech-to-speech path only provides one voice. We use the LFM3-Audio TTS to synthesize the S2S response in a different voice.

One caveat: the video generation stage is still GPU-intensive. For SoulX, using RTX 4090 or better GPUs.

cuDNN is a must for real-time performance. Without it inference is 2× slower and VRAM usage doubles.

\~850ms between submission a question and avatar speaks on an RTX 5090. ASR runs in parallel. As mentioned LFM3-Audio is called twice, which can be optimized. There’s also some overhead from the prototype UI. A native app with a more optimized implementation could further reduce latency.
```



document fetched 2026-11-16 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1xx2vi0/audiocpp_soulxflashhead_lfm3audio_owsm_v4
