aoiandroid / neutts-jp-150m-gguf

Model Card: neuTTS-JP-150m GGUF (multi-quantization)

Update (2026-05) : Q4_K_M alone is no longer recommended for this 150M model. Q4_K_M cannot emit model.config.eos_token_id = 2 (or tokenizer </s> = 65538 ) under greedy or temp=0.8 sampling and runs out to max_new_tokens = 1500 of low-FSQ noise. Use neutts-jp-150m-Q5_K_M.gguf (recommended) , Q8_0 , or F16 instead. Q4_K_M is kept for backwards compatibility only.

1. モデル詳細 (Model Details)

⚠️Attention⚠️ このリポジトリのモデルを全て実用レベルではありません。まだ開発段階です。実際の使用を想定したSwift runnerでのttsテストの結果によると、日本語の音声のように聞こえる変な声の出力、 または全ての言葉を正しく発音できていない不完全な日本語の音声が確認できました。これはユーザーによる手打ちうどんです。

• モデル名 : neuTTS-JP-150m (GGUF版)
• アーキテクチャ : LlamaForCausalLM (12 layers, hidden=512, ff=2048, 8 heads)
• パラメータ数 : 約 1.5 億 (150M)
• 言語 : 日本語 (Japanese)
• タスク : Text-to-Speech (TTS) — 音声トークン生成
• フォーマット : GGUF (llama.cpp 互換)
• コンテキスト長 : 4096
• 語彙サイズ : 99,111 (特殊トークン <audio_0> 〜 <audio_65535> , <SEP|LLM-jp> 等を含む)
• upstream : aoiandroid/neutts-jp-150m-float16
• 再変換 : tokenizer.json (Unigram + byte_fallback) → llama-style vocab、Q5_K_M / Q8_0 / F16

2. 推奨パラメータ

repetition_penalty: 1.1
temperature:        0.8
top_k:              40        # llama.cpp デフォルトと整合
max_new_tokens:     1500

EOS は 2 つあり 、両方を停止条件に含めること:

• model.config.eos_token_id = 2 ( <audio_2> ) ※ HF model.generate() のデフォルト停止
• tokenizer の </s> = 65538 ※ プロンプトのセパレータ。Q5/Q8 はこちらで自然停止

3. TTS パイプライン

1. トークナイズ:   入力テキスト → text_ids   (HF PreTrainedTokenizerFast)
2. 音声エンコード: 参照音声     → ref_codes  (NeuCodec encoder)
3. プロンプト:     ref_codes + [</s>=65538] + text_ids + [</s>=65538]
4. 推論:           本 GGUF (llama.cpp) で repetition_penalty=1.1 → top_k=40 → temp=0.8
5. デコード:       生成 ID から id < 65536 を抽出 → NeuCodec decoder → 24 kHz WAV

4. 使用例

4.1 llama.cpp CLI (Q5_K_M)

./llama.cpp/build/bin/llama-cli \
    -m ./neutts-jp-150m-Q5_K_M.gguf \
    -n 1500 \
    --no-display-prompt \
    --repetition-penalty 1.1 \
    --temp 0.8 \
    --top-k 40

4.2 iOS/macOS Swift (TranslateBlue 互換)

let model = LlamaCppModel()
try model.load(
    "/path/to/neutts-jp-150m-Q5_K_M.gguf",
    architectureOverride: "neutts_jp150m"
)
// ref_codes (NeuCodec encoder) と HF text_ids を Swift 側で構築
let generated = try model.generateNeuTTSJP150MCausalLMTokens(
    fromInputIds: prompt,
    eosTokenId: 2,            // model.config.eos_token_id
    secondaryEosTokenId: 65538, // tokenizer </s>
    maxNewTokens: 1500,
    minNewTokens: 8,
    repetitionPenalty: 1.1,
    vocabSize: 99111
)

5. 制限事項

• NeuCodec デコーダ必須 : 本モデルは音声トークン (Acoustic Tokens) のみ生成。24 kHz 波形化には neuphonic/neucodec 互換のデコーダパイプラインが必要。
• HF カスタムトークナイザ前処理 : 学習に用いた LlmJpTtsTokenizer をそのままアプリに同梱できない場合は、 tokenizer.json を PreTrainedTokenizerFast で読み込む経路で互換動作する。
• Q4_K_M は EOS が壊れる : Q4_K_M ロジットがフラットになり EOS ( 2 も 65538 も) を出さず、約 30 秒のノイズになる。 Q5_K_M 以上を使用すること 。
• メモリ : Q5_K_M で実行時 ~150 MB、Q8_0 で ~210 MB、F16 で ~360 MB が目安。