When generating a voice, Qwen3-TTS often adds a strange "word" or "words" to the end.

[ssechi2]

If you generate a voice with the same source model, but using a standard Python script, then there is no such problem, no additional "words" at the end of audio.
The problem exists on all versions of Qwen3-TTS.
Setting the recommended parameters for this model does not help. (--repeat_penalty 1.05 --temp 0.9 --top_p 1.0 --top_k 50 --max_new_tokens 8192)
Tested on backends: vulkan & cpu.
Cuda crash with "Eggml_cuda_compute_forward: IM2COL failed ECUDA error: invalid configuration argument "

[foldl]

TODO: check precision of talker.code_predict.