Qwen3-TTS distillation pushes compression limits

Hot take: TTS distillation is usually less forgiving than people expect, because you are compressing not just semantics but timing, prosody, speaker identity, and audio fidelity.

• –The ask is practical, but it signals a real research gap: shrinking a speech model without wrecking naturalness is a different problem than shrinking a text LLM.
• –“Garbage” output often means the student is missing acoustic alignment or prosody supervision, not just parameter capacity.
• –The useful advice here is likely to come from speech-specific distillation methods, paired audio-text training, and evaluation on intelligibility plus MOS-style quality, not just size reduction.
• –This is more of a technical help thread than a product launch story, so the main value is for people building or compressing speech models.