Sparkplugx1904/whisper-small-id

Whisper Small Model – Indonesian ASR

Model Description

This model is a fine-tuned version of openai/whisper-small for Automatic Speech Recognition (ASR) in Indonesian (id).
It supports transcription of Indonesian speech into text across various audio conditions, with performance and resource usage depending on the selected model size.

Intended Use

• Indonesian speech-to-text transcription
• Research and experimentation
• Educational and academic purposes
• Application development and benchmarking

Model variants (tiny, base, small, medium, large) differ in accuracy, speed, and hardware requirements. Users should select the size that best matches their constraints and objectives.

Limitations

• Transcription quality depends on audio clarity, speaker accent, and background noise
• Smaller variants may produce higher error rates on long or complex audio
• Larger variants require significantly more compute and memory
• Outputs should be reviewed before use in critical or high-risk applications

Training Data

This model was fine-tuned using Mozilla Common Voice v23.0 (Indonesian).
Common Voice is a publicly available, community-driven speech dataset released by Mozilla under a permissive license.
Dataset characteristics such as speaker diversity, recording quality, and utterance length may influence model behavior.

Evaluation

The model is typically evaluated using Word Error Rate (WER).
Evaluation results may vary depending on dataset, domain, audio conditions, and model size.

Training results

Step	Training Loss
100	0.897100
200	0.509400
300	0.234200
400	0.153100
500	0.068000
600	0.074100
700	0.029100
800	0.017800
900	0.013600
1000	0.007200
1100	0.004900
1200	0.003700
1300	0.001800
1400	0.001700
1500	0.001100
1550	0.001100