SpeechT5 fine-tuned on Dutch VoxPopuli

This model is a fine-tuned version of microsoft/speecht5_tts on the qmeeus/voxpopuli dataset. It achieves the following results on the evaluation set:

• Loss: 0.4977

Model description

Experimental Audio Course checkpoint fine-tuned for 1,000 optimizer updates on three hours of Dutch VoxPopuli audio.

Intended uses & limitations

This is a training exercise. Generated speech quality remains limited: the included sample's automatic transcription does not match the intended prompt. Accurate Dutch speech synthesis has not been established.

Training and evaluation data

Dutch qmeeus/voxpopuli audio was limited to three hours, with utterances between one and twenty seconds and short transcriptions. A fixed seed-42 split reserved ten percent for evaluation. Exact counts and the selected checkpoint are in training-summary.json.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 4
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 1000
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.46.3
• Pytorch 2.8.0+cu126
• Datasets 2.21.0
• Tokenizers 0.20.3