Modelpublic

pymlex/ipa-transcriptor-300M

sourceHugging Facegpl-3.0updated 16d agoView on Hugging Face

0likes365downloads

Model Card

IPA Transcriptor 300M

Fine-tuned `google/byt5-small` for English word to IPA transcription.

Task format:

text

ipa: analytical  ->  ˌænəˈlɪtɪkəl

Training data: English phonetic and syllable count dictionary, 125,925 word–IPA pairs after cleaning.

GitHub: pymlex/ipa-transcriptor-300M

Benchmark

Fine-tuned on NVIDIA L4, run colab_l4_bf16, beam search num_beams=4.

Metric	Validation	Test
`n_samples`	6296	6297
`loss`	0.1515	0.1478
`perplexity`	1.1636	1.1592
`token_accuracy`	0.7849	0.7858
`exact_match`	0.5982	0.6111
`char_accuracy`	0.8948	0.8969
`cer`	0.1066	0.1045
`bleu`	59.82	61.11

Dataset length distributions

[image]

Training loss

[image]

Inference

python

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch

model_id = "pymlex/ipa-transcriptor-300M"
source_prefix = "ipa: "

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
model.eval()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

def transcribe(word: str, num_beams: int = 4) -> str:
    source = f"{source_prefix}{word.strip().lower()}"
    encoded = tokenizer(source, return_tensors="pt", truncation=True, max_length=36).to(device)
    output_ids = model.generate(**encoded, max_new_tokens=56, num_beams=num_beams, early_stopping=True)
    return tokenizer.decode(output_ids[0], skip_special_tokens=True)

print(transcribe("analytical"))