basha3la2/whisper-base-de-v3

Whisper Base DE v3 - Alaa Albasha

This model is a fine-tuned version of basha3la2/whisper-base-de-v2 on the ASR German Mixed dataset.

It achieves the following results on the evaluation set:

• Loss: 0.2006
• Wer: 13.0467

Training History

This is the third round of fine-tuning Whisper Base for German ASR:

Round	Base Model	Dataset	Steps	WER
v1	openai/whisper-base	flozi00/common_voice_19_0_de-labeled	4,000	~15.6%
v2	basha3la2/whisper-base-de	flozi00/asr-german-mixed	15,000	14.16%
v3	basha3la2/whisper-base-de-v2	flozi00/asr-german-mixed	~10,000	13.05%

Training Note

Training for v3 was planned for 30,000 steps but was interrupted at approximately step 10,000 due to Kaggle's 12-hour session limit. The last checkpoint was saved and pushed to the Hub. Further training continues in v4.

Training Results

Training Loss	Step	Validation Loss	Wer
0.8029	5000	0.2134	13.7950
0.7785	10000	0.2006	13.0467

Training Hyperparameters

• learning_rate: 3e-06
• train_batch_size: 8
• eval_batch_size: 8
• gradient_accumulation_steps: 2 (effective batch size: 16)
• seed: 42
• optimizer: AdamW
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 200
• training_steps: 30,000 (interrupted at ~10,000)
• mixed_precision_training: FP16

Intended Use

German automatic speech recognition, deployed on Radxa Rock 5B+ via RKNN (NPU inference).

Framework versions

• Transformers 5.x
• PyTorch 2.x
• Datasets 5.x