skypro1111/gemma-3-270m-uk-verbalizer

Overview

This model, skypro1111/gemma-3-270m-uk-verbalizer, is a specialized Gemma 3 (270M parameters) fine-tuned for Ukrainian text normalization. Its primary function is to convert written Ukrainian text into its phonetic, spoken representation, making it ideal for Text-to-Speech (TTS) systems. It accurately verbalizes complex elements such as numbers, dates, times, currency, units, abbreviations, codes, phone numbers, IBANs, domains, emails, Roman numerals, and Latin insertions.

Key Capabilities

• Comprehensive Normalization: Handles a wide array of text types, ensuring correct pronunciation for TTS. Examples include "1111 386,40 грн" becoming "один мільйон сто одинадцять тисяч триста вісімдесят шість гривень сорок копійок".
• Contextual Inflection: Correctly inflects numerals based on context (e.g., "при двох тисячах восьмистах тридцяти відвідувачах") and distinguishes reading conventions (e.g., вітамін D as "вітамін де" vs. роз'єм типу C as "роз'єм типу сі").
• Optimized Performance: Achieves significantly high inference speeds, particularly with GGUF via llama.cpp, reaching 532 tokens/s on an RTX 3090 Ti, which is 20 times faster than transformers.
• Pruned Vocabulary: The original Gemma 3 vocabulary was surgically pruned from 262,144 to 38,651 tokens, reducing model size by 53% (from 270M to ~125M parameters) without retraining or quality loss, by focusing on Ukrainian-specific tokens.
• Preprocessing Steps: Includes essential deterministic preprocessing steps (digit_router and bignum_spacing) to handle specific numerical patterns that the model might otherwise misinterpret, ensuring higher accuracy.

Good for

• Ukrainian Text-to-Speech (TTS) Systems: Provides the crucial text verbalization step before speech synthesis.
• High-Performance Inference: Recommended for applications requiring fast processing of Ukrainian text, especially when deployed with llama.cpp.
• Resource-Constrained Environments: The pruned vocabulary and ONNX export options make it suitable for edge devices or CPU-based inference where llama.cpp might be inconvenient.

Limitations

• Transliteration: Performance on unseen English words is around 50%, as the model learned a list rather than a rule.
• Sentence-by-Sentence Processing: Designed for single sentences; long paragraphs should be split before input.
• Specific Domains: Weaker performance on Roman numerals in non-canonical forms, lists of multiple English words, bare email fragments, and scientific/technical texts.
• Hyphenated Ranges: Currently reads 5-7 млн as "п'ять — сім мільйонів" instead of "від п'яти до семи", which is a known issue to be addressed in future updates.