Qwen3-TTS · Issues· 55 open

[Commercial use clarification] May outputs from preset voice Sohee be used in monetized YouTube videos?

Finetuning loss misaligned with generation (double shift + leaked sub-talker hidden) — please merge #278

Any plan to release 25Hz tts model ？

Fine-tuned 1.7B voice: short/opening utterances unstable (wrong timbre, occasional gender flip) while long-form is stable

25Hz tokenizer: qkv_attention_manual never masks padded keys (bool masked_fill), short attention windows get wrong outputs

Multi-concurrency inference not supported

Verify evals on Papers with Code

finetuning/sft_12hz.py: checkpoint save is non-atomic (can silently leave base weights) and peaks at 2x model size in host RAM

Fine-tuning Qwen3-TTS-12Hz-1.7B-Base for Bengali: Foreign/Hindi Accent Bias and Lack of Naturalness

AMD MI300X - Qwen/Qwen3-TTS-12Hz-1.7B-Base - Voice Cloning Segmentation fault (core dumped)

Voice-clone (ICL) generation echoes the tail of the reference audio before the requested text

Feature request: native Italian preset voices (male + female)

Finetuning Base results in progressively faster speech with every epoch

The talker_hidden_states not obtained correctly

The audio generated for Japanese and Korean is incomplete.