lowhipa-large-sr (Sanna Related)

This Whisper-for-IPA (WhIPA) model adapter is a PEFT LoRA fine-tuned version of openai/whisper-large-v2 on a subset of:

• CommonVoice11 dataset (1k samples each from Greek, Maltese) with G2P-based IPA transcriptions
• Arabic Speech Corpus (https://en.arabicspeechcorpus.com) with custom IPA transcriptions transliterated from the provided Buckwalter transcriptions (1k samples) (https://doi.org/10.5281/zenodo.17111977)

Model description

For deployment and description, please refer to https://github.com/jshrdt/whipa.

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

tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-large-v2", task="transcribe")
tokenizer.add_special_tokens({"additional_special_tokens": ["<|ip|>"] + tokenizer.all_special_tokens})

base_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v2")
base_model.generation_config.lang_to_id["<|ip|>"] = tokenizer.convert_tokens_to_ids(["<|ip|>"])[0]
base_model.resize_token_embeddings(len(tokenizer))

whipa_model = PeftModel.from_pretrained(base_model, "jshrdt/lowhipa-large-sr")

whipa_model.generation_config.language = "<|ip|>"
whipa_model.generation_config.task = "transcribe"

whipa_processor = WhisperProcessor.from_pretrained("openai/whisper-large-v2", task="transcribe")

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

Training results

Training Loss	Epoch	Validation Loss
0.4344	2.0323	0.3692754805088043
0.1875	4.0645	0.3102695643901825
0.0717	6.0968	0.30600059032440186
0.0202	8.1290	0.32697898149490356
0.0101	10.1613	0.34040552377700806

Framework versions

• PEFT 0.15.1
• Transformers 4.48.3
• Pytorch 2.6.0+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0