Qwen3‑TTS + Remotion Skill

This skill automates:

• Creating a Python 3.12 Conda env and installing qwen-tts (0.6B/1.7B CustomVoice)
• Generating 9 speaker samples in English (or chosen language)
• Rendering a vertical 1080×1920 Remotion video that plays each sample with a title card
• Optional: Creating one‑off voice replies from any text using Qwen3‑TTS and returning a local WAV path

Prereqs

• macOS with Homebrew
• Conda/Miniforge (auto-installed if missing)
• Node 20+ (Volta auto-installed if missing)
• FFmpeg (system)

Quick Start

# 1) Prepare env (Conda + Python 3.12) and install qwen-tts
bash scripts/setup-python.sh

# 2) Generate audio samples (English text by default)
TEXT="This is a voice test." bash scripts/generate-audio.sh

# 3) Build Remotion project and render
bash scripts/render-remotion.sh

# Output
# - Audio: clawd/tmp/qwen_samples/*_EN.wav
# - Video: projects/qwen3-tts-demo/out/QwenDemo.mp4

Voice replies (send any text as a voice message)

Generate a single WAV from arbitrary text and chosen speaker:

TEXT="Your message here" SPEAKER=Ryan LANG=English bash scripts/voice-reply.sh
# -> prints /Users/yusuf/clawd/tmp/qwen_voice/voice-<ts>.wav

Toggle auto-voice replies for agents/tools that consult the flag:

bash scripts/enable-voice-replies.sh   # turn ON (use TTS by default)
bash scripts/disable-voice-replies.sh  # turn OFF (send text)

Wrapper that respects the toggle:

bash scripts/reply-or-voice.sh "Your message" Ryan English

This path can be sent directly in chat systems that accept file uploads.

Customization

• Change the text: TEXT="Your line here" bash scripts/generate-audio.sh
• Change language/speaker list inside scripts/generate-audio.py
• Adjust scene timing in src/QwenDemo.tsx (sceneDur)

Notes

• For higher quality or cloning, switch to 1.7B or Base models in generate-audio.py.
• On CPU, generation is slower; GPU with FlashAttention 2 is recommended.