Whisper Medium Fine-tuned for Spontaneous Persian Speech
This model is a fine-tuned version of openai/whisper-medium
for Persian automatic speech recognition (ASR).
It was adapted on the Spontaneous Persian Speech (SPS) dataset,
with a focus on spontaneous and conversational Persian.
Model description
Base model: openai/whisper-medium
Language: Persian (Farsi)
Task: Automatic Speech Recognition
Adaptation method: Full fine-tuning
Training duration: 2 epochs
Training data
The model was trained on speech segments extracted from the
Spontaneous Persian Speech (SPS) dataset.
For this training run:
Training segments: 16,409
Validation segments: 1,985
The SPS dataset itself is not distributed with this model.
Training configuration
Learning rate: 5e-6
Warm-up ratio: 0.05
Batch size: 1
Gradient accumulation steps: 64
Effective batch size: 64
Maximum target length: 160 tokens
Gradient checkpointing: enabled
Evaluation
The saved evaluation run contains 500 scored test segments.
Mean segment-level results:
Model
WER
CER
Whisper-medium baseline
2.368
1.857
Fine-tuned Whisper-medium
0.768
0.506
These values are mean segment-level error rates rather than
corpus-level aggregated WER/CER.
The fine-tuned model substantially improves over the corresponding
Whisper-medium baseline on this evaluation set.
A separate phonological analysis was also conducted to examine
how fine-tuning affected phone-substitution errors and their
phonological characteristics.
Usage
The model can be loaded directly with Hugging Face Transformers
and used to transcribe Persian audio.
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
model_id = "Behnooshn/whisper-medium-sps-persian"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id)
Intended use
This model is intended primarily for research on Persian ASR,
particularly spontaneous and conversational Persian speech.
Performance on other speech domains, recording conditions,
or Persian varieties has not been systematically evaluated.
Limitations
The model was fine-tuned on a limited spontaneous Persian dataset.
Its performance should therefore not be assumed to generalize
equally to all speakers, domains, dialects, or acoustic conditions.
Data availability
The SPS training data are not included in this repository.
