Update version of model using phonemes

README.md

@@ -3,91 +3,81 @@ license: apache-2.0
 base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
 tags:
 - lyrics
+- phonetics
+- g2p
 - katakana
+- english-to-phoneme
 - english-to-katakana
-- katakana-english
-- english2katakana
+- liaison
 - tinyllama
 ---
 
-# TinyLlama-1.1B-Katakana-Lyrics-Liaison
+# TinyLlama-1.1B-Phonetic-Liaison-Katakana-Generator
 
-This model is a fine-tuned version of `TinyLlama/TinyLlama-1.1B-Chat-v1.0` using LoRA. It is specifically designed to convert English lyrics and phrases into **Phonetic Katakana**, prioritizing real-world pronunciation, linking (liaison), and rhythm over literal dictionary spelling.
+This model is a fine-tuned version of `TinyLlama/TinyLlama-1.1B-Chat-v1.0` designed to predict **connected phoneme sequences** and **rhythm-optimized Katakana**. It focuses on capturing real-world auditory phenomena like liaison, reduction, and flapping.
 
-## 🌟 Concept: "The Training Wheels for English Rhythm"
+## 🌟 The Concept: "Phonetic Bridge for Natural Speech"
 
-As the creator, I generally believe that English should be learned without Katakana. However, for children and beginners, the "fear of the written word" often stops them from speaking entirely.
+Traditional G2P (Grapheme-to-Phoneme) converters often treat words in isolation. This model serves as a **Phonetic Bridge**, predicting how sounds change in continuous speech.
 
-This model was built to provide **"Supportive Katakana"** — not just a translation, but a phonetic guide that helps learners mimic the actual rhythm and flow of native speakers, serving as temporary "training wheels" until they are ready to rely solely on their ears.
+### For Global Developers (The "Connected Phonemes" Advantage)
+While the model outputs Katakana, its core intelligence lies in generating **Connected Phoneme Sequences (ARPAbet)**. 
+- **TTS Frontend:** Use the linked phoneme output to improve the prosody of your Text-to-Speech engines.
+- **ESL Tools:** Visualize for learners how "Take it" becomes `/t ey1 k ih1 t/` instead of two separate words.
 
-## ✨ Key Features
-
-* **Liaison & Linking:** Handles word connections naturally (e.g., `hold your` → `ホージョ`, `take it` → `テイキッ`).
-* **Silent Letters:** Trained to ignore silent consonants (e.g., `honest` → `オネス`, `hour` → `アワー`).
-* **Lyric-focused Reductions:** Strong support for informal contractions like `gonna`, `wanna`, and `gotta`.
-* **Complex Phonetics:** Specifically trained to handle difficult phonetic mappings like `Scarborough Fair` → `スカーブラフェア`.
+### For Japanese Learners ("The Training Wheels")
+I am a firm believer that English should ideally be learned through ears, not Katakana. However, beginners often face a "fear of the written word." 
+This model provides **"Supportive Katakana"**—not a translation, but a phonetic map that mimics native rhythm, acting as training wheels for the ear.
 
-## 📊 Comparison Examples
+## ✨ Key Features
 
-| English Phrase | Dictionary-style (Standard) | **This Model (Phonetic)** |
-| --- | --- | --- |
-| I wanna hold your hand | アイ ウォナ ホールド ユア ハンド | **アイワナホージョハン** |
-| I gotta be honest with you | アイ ガッタ ビー オネスト ... | **アイガラビーオネスウィズユー** |
-| Scarborough Fair | スカーバラ フェア | **スカーブラフェア** |
-| Take it anymore | テイク イット エニモア | **テイキッエニモー** |
+* **Connected Phonemes (ARPAbet):** Outputs the exact phonetic string including liaison (e.g., `a little bit` -> `AH0 L IH1 D AH0 L B IH1 T`).
+* **Liaison & Flapping:** Naturally handles `T` to `D` transformations and word-to-word connections.
+* **Silent Letters:** Intelligently ignores non-vocalized consonants.
+* **Modern ESL Approach:** Designed for high-speed inference on mobile devices (ready for GGUF/on-device PoC).
 
-## 🚀 How to Use
+## 📊 Comparison: Beyond Dictionary Rules
 
-To get the best results, use the following prompt format:
 
-```text
-英語を歌いやすいように、音のつながり（リエゾン）を考慮してカタカナに変換してください。
 
-英語: take it easy
-カタカナ: テイキッイージー
+| English Phrase | Dictionary Phonemes | **This Model (Linked Phonemes)** | **Supportive Katakana** |
+| --- | --- | --- | --- |
+| **A little bit** | `[AH0] [L IH1 T AH0 L] [B IH1 T]` | `AH0 L IH1 D AH0 L B IH1 T` | **アリロビッ** |
+| **Check it out** | `[CH EH1 K] [IH1 T] [AW1 T]` | `CH EH1 K IH1 T AW1 T` | **チェキラッ** |
+| **Middle of the night**| `[M IH1 D AH0 L] [AH1 V]...` | `M IH1 D AH0 L AH1 V DH AH0 N AY1 T`| **ミドロヴザナイッ** |
 
-英語: I wanna hold you
-カタカナ: アイワナホージュー
+## 🚀 Prompt Format
 
-英語: [Your English Phrase Here]
-カタカナ:
-```
+To extract both Katakana and the connected phoneme sequence, use the following format:
 
-### Example Code (Python / Transformers)
-
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftModel
-
-base_model_path = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
-lora_model_path = "YOUR_USERNAME/TinyLlama-1.1B-Katakana-Lyrics-Liaison"
+```text
+英語とその単語単位の音素から、リエゾンを考慮したカタカナと繋がった音素列を生成してください。
 
-tokenizer = AutoTokenizer.from_pretrained(base_model_path)
-model = AutoModelForCausalLM.from_pretrained(base_model_path)
-model = PeftModel.from_pretrained(model, lora_model_path)
+英語: take it easy
+単語音素: [T EY1 K] [IH1 T] [IY1 Z IY0]
+カタカナ: テイキットイージー
+繋がった音素: T EY1 K IH1 T IY1 Z IY0
 
-prompt = "英語を歌いやすいように、音のつながり（リエゾン）を考慮してカタカナに変換してください。\n\n英語: take it easy
-カタカナ: テイキッイージー\n\n英語: I wanna hold you\nカタカナ: アイワナホージュー\n\n英語: I love the way you lie\nカタカナ:"
-inputs = tokenizer(prompt, return_tensors="pt")
-outputs = model.generate(**inputs, max_new_tokens=50)
-print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+英語: {Your Phrase}
+単語音素: {Standard G2P Output}
+カタカナ: 
 
 ```
 
-## 🛠 Training Details
+## 🛠 Technical Specs & Dataset
 
-- **Dataset:** 1,200+ samples of **custom-curated phonetic pairs**. 
-- **Methodology:** Developed using a "human-in-the-loop" approach, focusing on capturing real-world auditory experiences rather than robotic dictionary rules.
-- **Fine-tuning Method:** LoRA (Low-Rank Adaptation)
-- **Base Model:** TinyLlama-1.1B-Chat-v1.0
+* **Dataset:** 1,200+ hand-curated pairs of English phrases and their auditory-correct phonetic mappings.
+* **Evaluation:** Currently being benchmarked against the `speechocean762` dataset for pronunciation scoring PoC.
+* **Architecture:** LoRA fine-tuning on TinyLlama 1.1B.
+* **Optimization:** Highly compatible with **GGUF** for ultra-lightweight mobile app integration (MFCC/DTW based evaluation).
 
-## ⚠️ Limitations
+## ⚠️ Limitations & Bias
 
-* **Model Size:** As a 1.1B model, it may occasionally hallucinate or misinterpret extremely long or rare technical terms.
-* **Dialect:** Primarily targets General American/Standard English pronunciation as heard in global pop music.
+* **Model Size:** 1.1B parameters. While fast, it may hallucinate on rare proper nouns.
+* **Accent:** Optimized for General American English (GenAm) commonly found in global pop music and media.
 
 ## 📜 License
 
-This model is licensed under the **Apache 2.0 License**, consistent with the base TinyLlama model.
+Apache 2.0
 
----
+```