kruokruo/wav2vec2-large-xlsr-53-kapampangan

Model Card for Wav2Vec2-Large-XLSR-53-Kapampangan

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 for Automatic Speech Recognition (ASR) in Kapampangan. It was trained on a custom dataset of 2,489 audio-transcription pairs and achieves competitive results on a held-out test split. The model supports transcription of Kapampangan speech, with partial support for English code-switching.

Model Details

Model Description

• —Developed by: Sean Almendral
• —Model type: Speech-to-text (ASR)
• —Languages (NLP): Kapampangan (kap), English (en)
• —License: Apache-2.0
• —Finetuned from model: facebook/wav2vec2-large-xlsr-53
• —Pipeline tag: automatic-speech-recognition

Model Sources

• —Repository: Hugging Face Model Card
• —Paper \[base model]: Wav2Vec2: Self-Supervised Learning of Speech Representations

Uses

Direct Use

• —Transcribing spoken Kapampangan into written text.
• —Educational and cultural preservation of Kapampangan.
• —Can be used as a baseline ASR model for further fine-tuning.

Downstream Use

• —Integration into translation pipelines (e.g., Kapampangan → English).
• —Voice-enabled chatbots, transcription tools, or accessibility apps.

Out-of-Scope Use

• —Not suited for medical, legal, or safety-critical transcription.
• —Performance may degrade with noisy audio or dialects outside the training data.

Bias, Risks, and Limitations

• —Dataset size is relatively small (2,489 clips), so coverage is limited.
• —May not generalize well to informal speech, background noise, or rare words.
• —English code-switching is partially supported but not fully accurate.
• —Risk of transcription bias due to imbalanced speaker representation in the dataset.

Recommendations

• —Users should validate outputs before using them in sensitive applications.
• —For production, additional fine-tuning with larger and more diverse Kapampangan datasets is recommended.

How to Get Started with the Model

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="kruokruo/wav2vec2-large-xlsr-53-kapampangan"
)

result = asr("sample_audio.wav")
print(result["text"])

Training Details

Training Data

• —Custom dataset of 2,489 audio-transcription pairs in Kapampangan.
• —Audio normalized to 16kHz mono WAV.
• —Preprocessing included lowercasing, punctuation removal, and character-based vocabulary building.

Training Procedure

• —Base model: facebook/wav2vec2-large-xlsr-53
• —Objective: Connectionist Temporal Classification (CTC)
• —Optimizer: AdamW
• —Learning rate: 1e-4
• —Epochs: 30
• —Train/Validation split: 80/20 (random seed=42)

Preprocessing

• —Audio: resampled to 16kHz mono WAV
• —Text: lowercased, punctuation removed, character-level tokenization

Training Hyperparameters

• —Precision: fp32
• —Batch size: dependent on GPU memory (typical 16–32)
• —Scheduler: Linear decay with warmup

Evaluation

Testing Data

• —20% held-out split from the same dataset (≈498 samples).

Metrics

• —WER (Word Error Rate): 0.3586
• —CER (Character Error Rate): 0.1259

Results

• —The model achieves strong performance for clean Kapampangan speech.
• —Maintains low character-level error (12.6%) and moderate word-level error (35.9%). ---

Environmental Impact

• —Hardware Type: NVIDIA L4 GPU (Google Colab Pro)
• —Training Duration: \[Fill in total hours used, if tracked]
• —Energy Consumed: \~514.06 Wh
• —Carbon Emissions: \~209.74 gCO2e
• —Cloud Provider: Google Colab Pro
• —Compute Region: Not disclosed by Google Colab

These numbers were estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute).

Technical Specifications

Model Architecture and Objective

• —Architecture: Wav2Vec2-Large-XLSR-53 (317M parameters, trained on 53 languages)
• —Objective: Fine-tuned with a CTC head for Kapampangan ASR

Compute Infrastructure

• —Framework: PyTorch + Hugging Face Transformers
• —Trainer: Hugging Face Trainer API

Citation

BibTeX:

@misc{wav2vec2-kapampangan,
  title={Wav2Vec2-Large-XLSR-53-Kapampangan: Automatic Speech Recognition Model},
  author={Sean Almendral},
  year={2025},
  howpublished={\url{https://huggingface.co/your-username/wav2vec2-large-xlsr-53-kapampangan}},
}

Model Card Authors

• —Sean Almendral

Model Card Contact

• —seansimonea@gmail.com