Qwen3-TTS-12Hz-1.7B Fine-tuned (French SIWIS)

Fine-tuned version of Qwen/Qwen3-TTS-12Hz-1.7B-Base on the French SIWIS dataset for improved French TTS.

Training Details

  • Base model: Qwen/Qwen3-TTS-12Hz-1.7B-Base
  • Training data: SIWIS French Speech Synthesis Database (~8,325 samples, 500 benchmark phrases excluded)
  • Training type: Full fine-tuning (speaker encoder frozen)
  • Best checkpoint: Epoch 6/10, val_loss=7.2436

Hyperparameters

Parameter Value
Epochs 6 (best of 10)
Batch size 4
Gradient accumulation 8
Effective batch size 32
Learning rate 2e-6
Warmup steps 500
Weight decay 0.01
LR scheduler Cosine
Precision bf16

Training Loss Progression

Epoch Train Loss Val Loss
1 12.89 9.13
2 8.15 7.72
3 7.53 7.43
4 7.35 7.31
5 7.27 7.26
6 7.23 7.24

Hardware

  • GPU: NVIDIA RTX A5500 (24GB)

Usage

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

tts = Qwen3TTSModel.from_pretrained(
    "Rcarvalo/qwentts",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

wavs, sr = tts.generate_custom_voice(
    text="Bonjour, comment allez-vous aujourd'hui?",
    speaker="siwis_french",
)
sf.write("output.wav", wavs[0], sr)

Baseline Benchmark (before fine-tuning, 500 SIWIS phrases)

Metric Value
WER (mean) 23.4%
WER (median) 14.3%
RTF (mean) 1.300

License

Apache 2.0 (same as base model)

Downloads last month
23
Safetensors
Model size
2B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Rcarvalo/qwenTTS

Finetuned
(61)
this model