.agents/memory/2026-01-26.md

2026-01-27

voice messaging setup completed

successfully implemented full voice message support for telegram:

tts (text-to-speech):

• qwen3-tts-12hz-1.7b-base model
• voice cloned from nicholai's alan rickman/snape impression
• reference audio: /mnt/work/clawdbot-voice/reference_snape_v2.wav
• running as systemd service on port 8765
• larger 1.7b model chosen over 0.6b for better british accent preservation

transcription (speech-to-text):

• using pywhispercpp (whisper.cpp python bindings)
• same setup as hyprwhspr
• model: base.en with 4 threads
• known issue: longer messages get truncated (only first segment captured)

workflow:

1. receive voice message → transcribe with whisper
2. respond with text
3. generate voice reply via tts service (port 8765)
4. send voice message back

all documented in VOICE-WORKFLOW.md

project location: /mnt/work/clawdbot-voice/

setup by: sonnet (with opencode sub-agent assistance)

tts service idle timeout added

implemented automatic vram management for qwen3-tts service:

changes:

• created tts_service_idle.py with idle timeout functionality
• model unloads after 120 seconds of inactivity (frees ~3.5GB VRAM)
• lazy loading: model only loads on first request, not at startup
• background monitor task checks idle status every 10 seconds
• updated systemd service to use new idle-aware version

configuration:

• TTS_IDLE_TIMEOUT=120 (configurable via environment variable)
• service still runs continuously, just unloads model when idle
• automatically reloads on next TTS request

benefit: nicholai can use comfyui without tts service consuming vram when not actively generating speech