metadata

language:
  - ckb
tags:
  - generated_from_trainer
datasets:
  - PawanKrd/asr-ckb
metrics:
  - wer
model-index:
  - name: ASR CKB
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: PawanKrd/asr-ckb
          type: PawanKrd/asr-ckb
        metrics:
          - name: Wer
            type: wer
            value: 44.878759772094874

Automatic Speech Recognition - CKB

This model is trained on the PawanKrd/asr-ckb dataset. It achieves the following results on the evaluation set:

Loss: 0.1310
Wer: 44.8788

Model description

This model is designed for automatic speech recognition (ASR) of the Central Kurdish language (Sorani dialect). It leverages a transformer-based architecture to transcribe spoken Kurdish into text. The model was trained using data that includes various speech samples representative of the language's phonetic diversity.

Intended uses & limitations

Intended Uses

Transcribing spoken Kurdish into text for applications such as subtitling, voice assistants, and transcription services.
Enhancing accessibility for Kurdish speakers by providing speech-to-text functionality in their native language.

Limitations

The model's performance may degrade with speakers who have strong accents or dialects not well-represented in the training data.
It may not perform well in noisy environments or with overlapping speech.
The Wer (Word Error Rate) of 44.8788 indicates room for improvement in accuracy.

Training and evaluation data

The model was trained on the PawanKrd/asr-ckb dataset, which consists of diverse speech recordings in Central Kurdish. The dataset includes a variety of speakers, both male and female, across different age groups and regions, providing a broad representation of the language.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 32
eval_batch_size: 16
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
training_steps: 15000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.253	0.1927	1000	0.3988	76.9180
0.2675	0.3854	2000	0.3212	65.9103
0.231	0.5780	3000	0.2816	61.2462
0.1703	0.7707	4000	0.2539	59.0665
0.1399	0.9634	5000	0.2321	55.0053
0.1671	1.1561	6000	0.2174	57.1154
0.1732	1.3487	7000	0.2026	54.5581
0.1258	1.5414	8000	0.1900	52.7660
0.1692	1.7341	9000	0.1817	57.1055
0.1854	1.9268	10000	0.1691	49.9702
0.2143	2.1195	11000	0.1588	48.6816
0.1562	2.3121	12000	0.1515	49.2083
0.0966	2.5048	13000	0.1445	47.7408
0.1071	2.6975	14000	0.1372	48.3868
0.0794	2.8902	15000	0.1310	44.8788

Framework versions

Transformers 4.42.0.dev0
Pytorch 2.3.0+cu121
Datasets 2.19.1
Tokenizers 0.19.1