Skip to content
 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

15 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Qwen3-TTS

https://github.com/project-felys/delta-me13

Data Preparation

python3 vendor/process_data.py \
  --jsonl 'cyrene/Chinese(PRC).jsonl' \
  --output-dir data/raw
python3 vendor/prepare_data.py \
  --device cuda:0 \
  --tokenizer_model_path Qwen/Qwen3-TTS-Tokenizer-12Hz \
  --input_jsonl 'data/raw/Chinese(PRC).jsonl' \
  --output_jsonl 'data/Chinese(PRC).jsonl'
python3 vendor/split_data.py \
  --jsonl 'data/Chinese(PRC).jsonl'

Training

Single run

nohup python3 vendor/sft_12hz.py \
  --init_model_path Qwen/Qwen3-TTS-12Hz-0.6B-Base \
  --output_model_path output \
  --train_jsonl 'data/Chinese(PRC).jsonl' \
  --speaker_name cyrene \
  --batch_size 8 \
  --gradient_accumulation_steps 2 \
  --num_epochs 4 \
  --lr 2.29e-5 \
  --min_lr_factor 0.44 \
  --lr_scheduler_type cosine \
  --warmup_ratio 0.05 \
  --save_last_ckpt_only \
  > "$(date +%s).log" 2>&1 &

Bayesian optimization sweep

nohup python3 vendor/bo_loop.py \
  --init_model_path Qwen/Qwen3-TTS-12Hz-0.6B-Base \
  --output_model_path output/bo/0 \
  --train_jsonl 'data/train/0/Chinese(PRC).jsonl' \
  --test_jsonl 'data/test/0/Chinese(PRC).jsonl' \
  --speaker_name cyrene \
  --seed 666 \
  > "$(date +%s).log" 2>&1 &

Inference

python3 vendor/infer.py \
  --model output/v4/ckpt-2 \
  --text 银河猫猫侠,我爱你~

Evaluation

python3 vendor/gen_audio.py \
  --test_jsonl 'data/test/0/Chinese(PRC).jsonl' \
  all \
  --model output/v4/ckpt-2 \
  --speaker_name cyrene
python3 vendor/gen_audio.py \
  --test_jsonl 'data/test/2/Chinese(PRC).jsonl' \
  baseline \
  --output_path output/baseline/2 \
  --ref_audio 'data/raw/ref.wav' \
  --ref_text '三千万篇相似而不相同的故事,听起来是不是很多呢?可我从来没有厌倦过,和你旅行的时候,又感动地重温了一遍呢!但故事的每一页,似乎都有些遗憾,想在这里加一段欢笑,想在那里填补上空白…多亏有你在,终于能为它们续写浪漫的结局了呀♪' \
  --language Auto
python3 vendor/eval_nn.py \
  --ground_truth_jsonl 'data/test/2/Chinese(PRC).jsonl' \
  --test_dir output/baseline/2

References

About

Qwen3-TTS is an open-source series of TTS models developed by the Qwen team at Alibaba Cloud, supporting stable, expressive, and streaming speech generation, free-form voice design, and vivid voice cloning.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages