miracle-leoma/qwen3-fine-tuned

Qwen3 Vocence Miner v2

This repository contains a Qwen3-TTS 12Hz 1.7B custom-voice checkpoint prepared for the Vocence PromptTTS subnet.

The model was fine-tuned from Qwen/Qwen3-TTS-12Hz-1.7B-Base on clean LibriTTS-R audiobook-style English speech. The target use case is long-form, clean narration for Vocence validator prompts.

Vocence Runtime

The repo includes the files expected by the canonical Vocence miner wrapper:

miner.py
chute_config.yml
vocence_config.yaml
config.json
model.safetensors
speech_tokenizer/

The miner contract is:

Miner(path_hf_repo)
Miner.warmup()
Miner.generate_wav(instruction: str, text: str) -> tuple[np.ndarray, int]

The /speak endpoint receives:

{
  "text": "exact transcript to speak",
  "instruction": "gender: male | pitch: mid | speed: normal | age_group: adult | emotion: neutral | tone: formal | accent: us"
}

This v2 checkpoint uses Qwen custom-voice generation with speaker:

vocence_narrator

Offline Vocence Eval Baseline

Scored on a held-out 100-sample LibriTTS-R validation set using the local Vocence evaluator from vocence/pipeline/evaluation.py.

win_rate: 49%
avg_score: 0.8886
script: 0.9879
naturalness: 0.56
gender: 0.81
speed: 0.995
emotion: 0.88
age_group: 1.00
pitch: 0.985
accent: 0.84
tone: 0.97

The main remaining weakness is naturalness. Script fidelity is strong.

Training Data

Training data was built from LibriTTS-R train-clean-100, chunked into 20-25 second same-speaker/same-chapter audiobook clips.

The training set used for this selected checkpoint was the top-20 speaker subset:

819 clips
1 epoch
lr: 2e-6
batch_size: 1

LibriTTS-R is derived from LibriTTS/LibriVox/Gutenberg materials and is distributed via OpenSLR under CC BY 4.0 terms.

Local Smoke Test

pip install qwen-tts soundfile numpy pyyaml

from pathlib import Path
from miner import Miner

m = Miner(Path("."))
wav, sr = m.generate_wav(
    instruction="gender: male | pitch: mid | speed: normal | age_group: adult | emotion: neutral | tone: formal | accent: us",
    text="The old house stood silent in the afternoon light, while the wind moved softly through the garden trees.",
)
print(wav.shape, sr)

Notes

This is an experimental Vocence miner checkpoint. It is optimized for clean English audiobook-style narration and exact transcript preservation, not broad expressive voice design.