Ormuri_ASR — Ormuri ASR Model

This model is a Low-Rank Adaptation (LoRA) fine-tuned version of Pashto Ghag (pashto-ghag-whisper-medium-asr, OpenAI Whisper Medium fine-tuned on Pashto) adapted specifically for Ormuri (oru), an endangered Indo-Iranian language spoken in Kaniguram (South Waziristan, Pakistan) and Logar (Afghanistan).

Model Highlights

• Base Model: Pashto Ghag (pashto-ghag-whisper-medium-asr)
• Target Language: Ormuri (oru)
• Architecture: WhisperForConditionalGeneration (Whisper Medium - 769M parameters)
• Fine-Tuning Method: PEFT / LoRA (r=32, alpha=64, modules: q_proj, v_proj, k_proj, out_proj, fc1, fc2)
• Training Epochs: ~1.62 (Checkpoint 600)
• Primary Metrics:
  • Word Error Rate (WER): 18.39% (down from 84.86% zero-shot baseline — -66.47 pp / 78.3% error reduction)
  • Character Error Rate (CER): 5.65% (down from 37.08% zero-shot baseline — -31.43 pp / 84.7% error reduction)

Benchmark Evaluation Results

Model Version	Target Language	WER (%)	CER (%)	Delta WER vs Baseline
Zero-Shot Pashto Ghag Base	Ormuri (oru)	84.86%	37.08%	Baseline
Fine-Tuned Checkpoint 100	Ormuri (oru)	36.54%	12.89%	-48.32 pp
Fine-Tuned Checkpoint 200	Ormuri (oru)	25.78%	8.35%	-59.08 pp
Fine-Tuned Checkpoint 300	Ormuri (oru)	23.18%	7.29%	-61.68 pp
Fine-Tuned Checkpoint 400	Ormuri (oru)	20.44%	6.41%	-64.42 pp
Fine-Tuned Checkpoint 600 (This Model)	Ormuri (oru)	18.39%	5.65%	-66.47 pp (Best)

Usage & Inference

Pipeline Inference (Easiest)

import torch
from transformers import pipeline

# Initialize pipeline
asr_pipeline = pipeline(
    "automatic-speech-recognition",
    model="burkialisher5/Ormuri_ASR",
    device="cuda" if torch.cuda.is_available() else "cpu",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
)

# Transcribe Ormuri audio file
result = asr_pipeline("path/to/ormuri_sample.wav")
print("Ormuri Transcription:", result["text"])

Manual Model Inference

import torch
import librosa
from transformers import WhisperProcessor, WhisperForConditionalGeneration

device = "cuda" if torch.cuda.is_available() else "cpu"

processor = WhisperProcessor.from_pretrained("burkialisher5/Ormuri_ASR")
model = WhisperForConditionalGeneration.from_pretrained(
    "burkialisher5/Ormuri_ASR",
    torch_dtype=torch.float16 if device == "cuda" else torch.float32
).to(device)

# Load audio resampled to 16kHz
audio, sr = librosa.load("path/to/ormuri_sample.wav", sr=16000)

input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features.to(device)
if device == "cuda":
    input_features = input_features.to(torch.float16)

forced_decoder_ids = processor.get_decoder_prompt_ids(language="pashto", task="transcribe")

with torch.no_grad():
    predicted_ids = model.generate(input_features, forced_decoder_ids=forced_decoder_ids, max_new_tokens=225)

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print("Ormuri Transcription:", transcription)

Dataset & Training Specs

• Corpus: Ormuri TTS & Speech Dataset (5,922 training audio clips, 705 validation clips, 706 test clips).
• Sampling Rate: 16,000 Hz.
• Hardware: NVIDIA GPU (FP16 mixed precision).
• Optimizer: AdamW (lr=3e-4).

Citation & Acknowledgements

• Pashto Ghag: Pre-trained Pashto ASR model (pashto-ghag-whisper-medium-asr).
• OpenAI Whisper: Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision.