pyon0024/tinyllama-katakana-converter

TinyLlama-1.1B-Phonetic-Liaison-Katakana-Generator Overview

This model, developed by pyon0024, is a specialized fine-tune of TinyLlama/TinyLlama-1.1B-Chat-v1.0 with 1.1 billion parameters. Its core function is to predict connected phoneme sequences (ARPAbet) and rhythm-optimized Katakana from English phrases. Unlike traditional Grapheme-to-Phoneme (G2P) converters that process words in isolation, this model focuses on how sounds change in continuous speech, handling phenomena like liaison, reduction, and flapping.

Key Capabilities

• Connected Phonemes (ARPAbet): Generates precise phonetic strings that account for word-to-word connections (e.g., "a little bit" becomes AH0 L IH1 D AH0 L B IH1 T).
• Liaison & Flapping: Accurately models sound changes such as 'T' to 'D' transformations and inter-word connections.
• Silent Letters: Intelligently identifies and ignores non-vocalized consonants.
• Supportive Katakana: Provides a phonetic map in Katakana that mimics native English rhythm, serving as a learning aid.
• High-Speed Inference: Optimized for mobile deployment, with strong compatibility for GGUF for on-device applications.

Good For

• TTS Frontend Development: Enhancing the prosody and naturalness of Text-to-Speech engines by providing linked phoneme outputs.
• ESL Tools: Visualizing phonetic changes in continuous speech for English language learners.
• Japanese Learners: Offering a "Phonetic Bridge" to understand English pronunciation and rhythm through Katakana, acting as "training wheels" for auditory learning.

Limitations

• Model Size: Due to its 1.1B parameter count, it may occasionally hallucinate on rare proper nouns.
• Accent: Primarily optimized for General American English (GenAm).