hailin
794ecbedc9
config: back to sherpa TTS, Qwen3 0.6B still too slow on 3090
2026-04-07 23:51:24 -07:00
hailin
912b54547a
config: switch to Qwen3-TTS 0.6B for faster inference
2026-04-07 08:55:08 -07:00
hailin
2ec0dcf401
config: switch back to sherpa TTS for better latency
2026-04-07 03:33:14 -07:00
hailin
d5eb5d6a38
fix: revert to single GPU, multi-GPU breaks TTS
2026-04-07 03:31:56 -07:00
hailin
17923f3bdc
feat: TTS on 2 GPUs (cuda:2,cuda:3) for faster inference
2026-04-07 03:28:16 -07:00
hailin
78bc3f71c0
fix: disable end_prompt, filter more thinking patterns from antaf
...
end_prompt was being sent to antaf as user message, causing LLM
to output its thinking process. Disabled end_prompt and added
more thinking patterns to filter.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 03:09:04 -07:00
hailin
ad2e91df8f
config: TTS speaker uncle_fu
2026-04-07 02:47:32 -07:00
hailin
e37a65ba1d
fix: TTS use cuda:2 (0 occupied)
2026-04-07 02:25:02 -07:00
hailin
d3fd9cc391
feat: add Qwen3-TTS CustomVoice GPU provider, switch TTS
...
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 02:21:46 -07:00
hailin
6707351540
config: remove all system prompts, let antaf use its own persona
...
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 02:01:15 -07:00
hailin
58c67de338
config: use Debian public IP for antaf bridge
...
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 01:47:38 -07:00
hailin
56042fddf7
config: switch LLM to antaf (text bridge)
...
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 01:41:34 -07:00
hailin
c55a7a010b
config: rename 阿福 to 小虎, wake word 小虎小虎
...
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 01:31:39 -07:00
hailin
30c95c3d2b
config: add 你好小智 to wakeup words (switch back to WakeNet)
...
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 00:38:11 -07:00
hailin
a3a5f749fb
config: add afu health AI agent prompts and config
...
- Role: 阿福, family doctor for elderly health consulting
- Prompt template: warm, patient, plain language, no jargon
- Wake words: 阿福阿福, 你好阿福
- ASR: Qwen3-ASR-1.7B (GPU cuda:1)
- TTS: sherpa-onnx local
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-07 00:08:40 -07:00
hailin