An Inverse Text Normalization (ITN) transformer model fine-tuned to convert unformatted, raw Latin Automatic Speech Recognition (ASR) outputs into fully formatted, classical Latin text. It simultaneously restores capitalization and trailing punctuation using a 14-class composite sequence labeling schema.

📌 Quick Links

Live Demo: Gradio Interface
Source Code: GitHub Repository
Base Model: latincy/latin-bert
Dataset: njand/latin-asr-post-processing-dataset

🛠️ Pipeline Architecture & Preprocessing

This model is intended to be used directly downstream of the acoustic model njand/wav2vec2-xls-r-latin.

Because raw ASR models emit stream-of-consciousness text (lowercased, space-separated, and unpunctuated), the text must pass through an input normalization pipeline before being fed into this model for casing and punctuation restoration.

text

+-----------------------+     +-------------------------------+     +-------------------------------+
|  Raw Audio Waveform   | --> |  njand/wav2vec2-xls-r-latin   | --> | Preprocessing & Normalization |
+-----------------------+     +-------------------------------+     +-------------------------------+
                                                                                    |
                                                                                    v
+-----------------------+     +-------------------------------+     +-------------------------------+
| Formatted Text Output | <-- |   Latin ASR Post-Processor    | <-- |  Custom CLTK Tokenization     |
+-----------------------+     +-------------------------------+     +-------------------------------+

Input Preprocessing Requirements

To prepare raw transcript outputs for inference, apply the following sequence of transformations:

Macron Stripping: Remove all vowel length diacritics (e.g., ā, ē, ī, ō, ū, ȳ → a, e, i, o, u, y).
Orthographic Standardization: Standardize consonant/vowel variants:
- Convert j → i and v → u.
- Handle orthographic exceptions (e.g., ejicio → eicio).
Custom CLTK Word Tokenization: Run the normalized string through a version of the CLTK (Classical Language Toolkit v0) Latin word tokenizer.

Note: Because official CLTK v0 tokenization scripts are unmaintained, a bespoke implementation of the tokenizer was executed dynamically during training preprocessing rather than being pre-applied to the static dataset.

🚀 Quickstart & Inference Utility

Below is a complete Python script demonstrating how to prepare raw ASR output and run inference using the post-processing pipeline.

python

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

PUNCT_MAP = {
    "NONE": "",
    "COMMA": ",",
    "PERIOD": ".",
    "SEMICOLON": ";",
    "COLON": ":",
    "QUESTION": "?",
    "EXCLAMATION": "!",
}

def format_token(word: str, tag: str) -> str:
    """Applies composite ITN tag (e.g., 'TITLE_COMMA') to a word token."""
    parts = tag.split("_")
    if len(parts) != 2:
        return word

    casing, punct = parts[0], parts[1]

    if casing == "TITLE":
        word = word.capitalize()
    elif casing == "LOWER":
        word = word.lower()

    return f"{word}{PUNCT_MAP.get(punct, '')}"

def restore_latin_text(pipe, raw_text: str) -> str:
    """Runs inference and reconstructs formatted Latin text."""
    predictions = pipe(raw_text, aggregation_strategy="first")
    formatted_words = [
        format_token(pred["word"].strip(" "), pred["entity_group"])
        for pred in predictions
    ]
    return " ".join(formatted_words)

# 1. Load pipeline
model_id = "njand/latin-asr-postprocessor"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

itn_pipe = pipeline("token-classification", model=model, tokenizer=tokenizer)

# 2. Test reconstruction with preprocessed ASR output
raw_asr_input = "gallia est omnis divisa in partes tres quarum unam incolunt belgae"
print(restore_latin_text(itn_pipe, raw_asr_input))
# Output: "Gallia est omnis divisa in partes tres, quarum unam incolunt Belgae."

🏷️ Composite Label Schema

Target labels utilize a 14-class composite sequence schema that pairs Casing state with Trailing Punctuation state:

$$\text{Label} = \text{Casing} \times \text{Punctuation}$$

Casing Tags (2): LOWER, TITLE
Punctuation Tags (7): NONE, COMMA, PERIOD, SEMICOLON, COLON, QUESTION, EXCLAMATION

⚙️ Model Variants & Optimization

To facilitate production deployment on CPU-based infrastructure, this repository provides the model in three formats:

Format	Precision	File Size	Latency (P50)	Recommended Use Case
PyTorch	FP32	443 MB	-	Training, fine-tuning, and PyTorch pipelines
ONNX	FP32	443 MB	21.5 ms	Production (Maximum Accuracy)
ONNX Quantized	INT8	188 MB	9.7 ms	Low-Latency & Edge CPU

Performance vs. Precision Trade-off: While INT8 dynamic quantization yields a 2.2× speedup (P50) and cuts RAM usage by 57.5%, top-line accuracy (92.16% → 91.03%) hides a severe drop in macro performance:

Macro F1 Collapse: Drops from 65.19% to 50.00%. Dynamic weight quantization compresses logit decision boundaries for rare token tags.

Punctuation Degradation: Punctuation F1 falls 10.66 percentage points (69.09% → 58.43%), causing increased missing or misclassified commas, colons, and sentence boundaries.

Casing Stability: Capitalization F1 remains mostly intact (91.71% → 89.38%).

Recommendation: Use ONNX FP32 for production pipelines where text formatting and punctuation precision are critical. Use ONNX INT8 in latency-critical environments where speed and memory constraints outweigh exact punctuation recovery.

📊 Benchmarks & Performance (Epoch 7 - Best Checkpoint)

Evaluated on a 95/5 train/holdout split across diverse Classical Latin literary and historical corpora.

Overall Summary Metrics

Metric	Score
Overall Accuracy	92.16%
Macro F1	0.6523
Precision	62.28%
Recall	69.43%
Validation Loss	0.2592

🎯 Sub-Task Breakdown

Task	Accuracy	F1 Score
Casing Restoration	98.35%	0.9171
Punctuation Insertion	93.67%	0.6908

📈 Training Progression

The model was trained over 9 epochs fine-tuning latincy/latin-bert. Model weights from Epoch 7 were selected based on optimal overall F1.

Epoch	Train Loss	Val Loss	Overall F1	Overall Acc	Casing Acc	Punct Acc
1	0.6284	0.2883	0.6097	91.05%	98.08%	92.79%
2	0.5617	0.2701	0.6275	91.60%	98.19%	93.26%
3	0.5150	0.2618	0.6388	91.91%	98.27%	93.50%
4	0.4855	0.2607	0.6455	92.04%	98.30%	93.60%
5	0.4652	0.2579	0.6440	92.12%	98.33%	93.65%
6	0.4502	0.2582	0.6462	92.15%	98.34%	93.66%
7	0.4351	0.2592	0.6523	92.16%	98.35%	93.67%
8	0.4262	0.2598	0.6495	92.24%	98.36%	93.75%
9	0.4172	0.2601	0.6507	92.21%	98.37%	93.71%

⚡ Hardware & Environmental Footprint

Hardware Infrastructure: NVIDIA L4 GPU via Modal
Training Time: 3.82 hours
Estimated Carbon Emissions: 0.1797 kg CO₂eq