uzair0/Katib-ASR

🪶 Katib ASR: State-of-the-Art Pashto Speech Recognition

Listening to the voices that the AI boom forgot.

Katib ASR is the most capable open-source Automatic Speech Recognition (ASR) model for the Pashto language (پښتو). Built on top of Whisper Large v3 and fine-tuned on the largest curated Pashto speech corpus assembled to date, Katib ASR brings real-time, highly accurate speech-to-text capabilities to millions of Pashto speakers.

---

🩸 The Story Behind Katib

Building state-of-the-art AI usually takes massive corporate research labs, entire teams of engineers, and unlimited compute. Katib ASR was built entirely solo.

The generative AI revolution is moving fast, but regional languages are being left behind. While developing voice-activated AI agents for medical clinics in Pakistan, the bottleneck became painfully clear: there was no reliable, high-fidelity transcription for Pashto.

Training an ASR model for a low-resource language is a massive grind. It meant hunting down scarce, fragmented audio datasets, writing custom text normalizers to fix broken Arabic-script transcriptions, and maximizing A100 GPU compute to ensure the architecture could handle the complex phonetics of native Pashto speakers.

Katib ASR is the result of that struggle — a dedicated, open-source model designed to give Pashto speakers a voice in the digital age.

---

---

🏆 Model Architecture & Performance

This is not a generic multilingual model. Katib ASR is a dedicated, purpose-built Pashto ASR system — the only model of its kind at this scale.

Feature	Detail
🧠 Base Model	Whisper Large v3 (1.55B parameters)
🗣️ Language	Pashto (پښتو) — Afghan & Pakistani dialects
⚡ Hardware	NVIDIA A100 80GB
🔢 WER	28.23% — best published result for open Pashto ASR

Evaluation Results

Evaluated on a held-out Pashto test set not seen during training:

Metric	Score
Word Error Rate (WER)	28.23%
Evaluation Loss	0.3011

💡 For context: The base whisper-large-v3 model — with no Pashto fine-tuning — produces largely garbled or Arabic-language output on Pashto audio. Katib ASR delivers coherent, structured transcriptions where the base model fails entirely.

---

📚 Datasets & Text Normalization

Katib ASR was trained on a multi-source, multi-dialect Pashto speech corpus carefully assembled and preprocessed from:

• Common Voice Pashto 24
• FLEURS Pashto
• A Custom Curated Pashto Corpus of in-house recordings

Custom Pashto Text Normalization

A key contribution of this model is a dedicated Pashto text normalization pipeline applied consistently to both training labels and inference output. It handles script variant inconsistencies across sources:

• Arabic Kaf (ك) → Pashto Kaf (ک)
• ݢ / گ → Pashto Gaf (ګ)
• Arabic Yey / Alef Maqsura variants → Pashto Yey (ی)
• All non-Arabic-script noise and punctuation removed

This ensures the model produces clean, standardized Pashto script regardless of the source audio's original transcription.

---

🚀 Quick Start

Using the Pipeline (Recommended)

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="uzair0/Katib-ASR",
    torch_dtype="auto",
    device="cuda",
    chunk_length_s=30,
)

result = asr("pashto_audio.wav")
print(result["text"])
# Example output: "زه غواړم چی ښار ته لاړ کړم"

Direct Model Loading

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

processor = WhisperProcessor.from_pretrained("uzair0/Katib-ASR")
model = WhisperForConditionalGeneration.from_pretrained(
    "uzair0/Katib-ASR",
    torch_dtype=torch.bfloat16
).to("cuda")

model.generation_config.language = "pashto"
model.generation_config.task = "transcribe"
model.generation_config.forced_decoder_ids = None
model.generation_config.suppress_tokens = []

---

⚙️ Training Configuration

Parameter	Value
Base model	whisper-large-v3
Precision	bfloat16 + TF32
Effective batch size	128 (64 × 2 grad accumulation)
Learning rate	1e-5 (linear schedule)
Warmup steps	92
Epochs	3
Optimizer	AdamW Fused
Gradient checkpointing	✅ Enabled

---

👨‍💻 Author & Citation

Built from the ground up by Muhammad Uzair at the University of Peshawar.

If you use Katib ASR in your research or applications, please consider citing it:

@misc{katibasr2026,
  title     = {Katib ASR: State-of-the-Art Pashto Automatic Speech Recognition},
  author    = {Muhammad Uzair},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/uzair0/Katib-ASR}
}

---

Built with ❤️ for the Pashto-speaking world.