balochi-whisper-medium

Balochi Whisper Medium Arabic-script LoRA

Fine-tuned PEFT LoRA adapter for Balochi ASR in Arabic script.

Base model

openai/whisper-medium

Output script

Balochi Arabic script.

Training summary

• Base model: openai/whisper-medium
• Adapter: PEFT LoRA
• Output script: Balochi Arabic script
• LoRA rank: 8
• GPU: NVIDIA Tesla T4
• Max steps: 800
• Final eval loss: approximately 1.213
• Combined data sources: BalASR + Common Voice / Balochi ASR dataset

Important inference rule

Balochi is not a built-in Whisper language token. Do not force English or translation.

model.config.forced_decoder_ids = None
model.generation_config.forced_decoder_ids = None

Loading example

import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

base_model_id = "openai/whisper-medium"
adapter_id = "shehzadkhalid04/balochi-whisper-medium"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if torch.cuda.is_available() else torch.float32

processor = WhisperProcessor.from_pretrained(adapter_id, subfolder="processor")

base_model = WhisperForConditionalGeneration.from_pretrained(
    base_model_id,
    torch_dtype=dtype,
    low_cpu_mem_usage=True,
    use_safetensors=True,
).to(device)

model = PeftModel.from_pretrained(base_model, adapter_id).to(device)
model.eval()

model.config.forced_decoder_ids = None
model.generation_config.forced_decoder_ids = None