Whisper-tiny-fa is an automatic speech recognition model specifically adapted for Persian (Farsi) speech. It builds upon OpenAI’s Whisper-tiny architecture, leveraging transfer learning to specialize in transcribing Persian audio. The model is suitable for converting spoken Persian audio into text, enabling applications such as voice assistants, captioning, and speech-driven user interfaces.

Base model: openai/whisper-tiny
Fine-tuned on: Common Voice 17.0 Persian subset
Languages supported: Persian (Farsi)
Model type: Encoder-decoder transformer (speech-to-text)

Intended uses & limitations

Intended uses:

Transcribing Persian (Farsi) speech to text from audio files or microphone input. Voice-controlled applications and speech interfaces for Persian speakers. Generating subtitles and closed captions in Persian for audio/video content.

Limitations:

The model is fine-tuned for Persian and may perform poorly on other languages. Performance may degrade with low-quality or noisy audio, accents, or dialects not well represented in the training data. Not suitable for real-time applications with strict latency constraints due to model size and processing requirements.

Training and evaluation data

Dataset: Common Voice 17.0 (Persian subset) Data split: Training, validation, and test splits provided by Common Voice were used. Preprocessing: Audio files were resampled to 16kHz and normalized. Transcripts were cleaned and normalized to standard Persian orthography.

Training procedure

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

Test results

Best test WER (Word Error Rate): 0.915
Best test CER (Character Error Rate): 0.428

Usage

python

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "aictsharif/whisper-tiny-fa"
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

result = pipe('sample.mp3')
print(result["text"])

Framework versions

Transformers 4.51.3
Pytorch 2.5.1+cu124
Datasets 3.5.0
Tokenizers 0.21.1