jasmine-yomi-2b
accurate Japanese G2P converter
2 min readupdated
jasmine-yomi-2b by pakutoma handles G2P (grapheme-to-phoneme conversion, translating mixed kanji and kana text into context-aware hiragana readings). Built on Qwen/Qwen3.5-2B (Qwen3_5ForCausalLM architecture with 1.88B parameters), it underwent full fine-tuning (Full SFT) on the ReazonSpeech v2 and CC-100 corpora. Unlike the base model, it is narrowly specialized in accurately determining character readings within Japanese sentences. It is useful for developers building voice assistants, text-to-speech (TTS) systems, and Japanese text processing tools.
Key features
- High accuracy of 99.05% on the strict JKYB-Parakeet benchmark.
- Built on the compact Qwen/Qwen3.5-2B LLM.
- Custom grammar handling: keeps particles like "ha" and "he" in their original spelling rather than phonetic form.
- Ready-to-use GGUF versions available for running via llama.cpp.
What you can do
- Integrate the model into a text-to-speech (TTS) pipeline for accurate kanji pronunciation.
- Process and normalize Japanese text prior to linguistic analysis.
- Run local inference using transformers or vLLM.
Example
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "pakutoma/jasmine-yomi-2b"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
model.eval()
text = "今日は良い天気ですね。"
inputs = tok(text + "", return_tensors="pt", add_special_tokens=False).to(model.device)
with torch.inference_mode():
out = model.generate(**inputs, max_new_tokens=384, do_sample=False)
yomi = tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(yomi) # きょうはよいてんきですね。
Read more
Details• Accuracy: 99.05% (13,407 out of 13,536 sentences)
• Relaxed Accuracy: 99.10%
• Target Kana-CER (raw / @1): 0.89% / 0.80%
• Architecture: 24 layers, hidden_size=2048, bfloat16 weights
• License: Apache-2.0










Comments