surafelabebe/whisper-small-am

Whisper Small Amharic

This model is a fine-tuned version of openai/whisper-small on the Common Voice 17.0 and surafelabebe/fleurs_am (a subset of google/fleurs) datasets. It achieves the following results on the evaluation set:

• Loss: 0.4352
• Wer: 50.9657

Model description

The model was trained for 10 hours on T4 GPU. Training results indicate potential overfitting. Future improvements will focus on mitigating this by incorporating a larger dataset, extended training epochs, and dropout regularization.

Usage

from transformers import pipeline

pipe = pipeline(model="surafelabebe/whisper-small-am")

text = pipe("sample.wav")["text"]  # change to "your audio file name"

print(text)

Input	Output Transcript
	አቶ ቦጋለ መብራቱ ወይዘሮ ውድነሽ በታሙም ባገቡ በሁለተኛው አመት መጫረሻ ወንድሪክ ሰውለደላቸውን
	ከሰብ ለሚሁን ከወይዘሮ ትሩ ወይም ከአብት ሺሰር ጋር ልዩሩ ጉዳይ ኖሮት አይደለም

Training procedure

The fine-tuning process followed a similar procedure to that described in this blog post.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• 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: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0108	9.6154	1000	0.3446	54.9759
0.0009	19.2308	2000	0.4052	51.7570
0.0001	28.8462	3000	0.4277	50.9388
0.0001	38.4615	4000	0.4352	50.9657

Framework versions

• Transformers 4.49.0
• Pytorch 2.5.1+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0