Thanks for the fix, this was a solid root cause analysis.

I found one thing after merging though. All Qwen3-TTS variants (Base, CustomVoice, VoiceDesign) share the same generate(voice=, instruct=) signature, so checking for "voice" in gen_params is always true for every Qwen3-TTS model. This means the elif "instruct" branch never runs for VoiceDesign either.

In practice, if someone sends voice to a VoiceDesign model, it goes to the voice kwarg (which VoiceDesign ignores) and instruct stays empty, so the model raises a 500.

I added two follow-up commits on main:

1. Changed AudioSpeechRequest.voice default from "default" to None. No model recognized "default" as a valid speaker anyway, and None lets models use their own defaults (Kokoro defaults to "af_heart", etc).

2. Added an instructions field to AudioSpeechRequest. This maps directly to the instruct kwarg in generate(), so VoiceDesign users can pass voice descriptions through instructions instead of voice. It also means CustomVoice can now receive both a speaker name (voice) and a voice description (instructions) at the same time.

@jundot Thanks for the follow-up — that's a much cleaner solution. I hadn't considered that all Qwen3-TTS variants share the same signature, so the voice-first check would break VoiceDesign. Adding a separate instructions field makes the intent explicit at the API level instead of trying to infer it from the model signature. Good call on the None default too.