diff --git a/crates/voicevox_core/src/__internal/doctest_fixtures.rs b/crates/voicevox_core/src/__internal/doctest_fixtures.rs
index dd088b218..c9029079c 100644
--- a/crates/voicevox_core/src/__internal/doctest_fixtures.rs
+++ b/crates/voicevox_core/src/__internal/doctest_fixtures.rs
@@ -1,12 +1,12 @@
-use std::{path::Path, sync::Arc};
+use std::path::Path;
 
 use crate::{AccelerationMode, InitializeOptions, OpenJtalk, Synthesizer, VoiceModel};
 
 pub async fn synthesizer_with_sample_voice_model(
     open_jtalk_dic_dir: impl AsRef<Path>,
-) -> anyhow::Result<Synthesizer> {
+) -> anyhow::Result<Synthesizer<OpenJtalk>> {
     let syntesizer = Synthesizer::new(
-        Arc::new(OpenJtalk::new(open_jtalk_dic_dir).await.unwrap()),
+        OpenJtalk::new(open_jtalk_dic_dir).await?,
         &InitializeOptions {
             acceleration_mode: AccelerationMode::Cpu,
             ..Default::default()
diff --git a/crates/voicevox_core/src/engine/open_jtalk.rs b/crates/voicevox_core/src/engine/open_jtalk.rs
index e45f9a5fa..1cffe6757 100644
--- a/crates/voicevox_core/src/engine/open_jtalk.rs
+++ b/crates/voicevox_core/src/engine/open_jtalk.rs
@@ -1,9 +1,6 @@
 use std::io::Write;
 use std::sync::Arc;
-use std::{
-    path::{Path, PathBuf},
-    sync::Mutex,
-};
+use std::{path::Path, sync::Mutex};
 
 use anyhow::anyhow;
 use tempfile::NamedTempFile;
@@ -21,9 +18,10 @@ pub(crate) struct OpenjtalkFunctionError {
 }
 
 /// テキスト解析器としてのOpen JTalk。
+#[derive(Clone)]
 pub struct OpenJtalk {
     resources: Arc<Mutex<Resources>>,
-    dict_dir: Option<PathBuf>,
+    dict_dir: Arc<String>, // FIXME: `camino::Utf8PathBuf`にする
 }
 
 struct Resources {
@@ -36,46 +34,41 @@ struct Resources {
 unsafe impl Send for Resources {}
 
 impl OpenJtalk {
-    // FIXME: この関数は廃止し、`Synthesizer`は`Option<OpenJtalk>`という形でこの構造体を持つ
-    pub fn new_without_dic() -> Self {
-        Self {
-            resources: Mutex::new(Resources {
+    pub async fn new(open_jtalk_dict_dir: impl AsRef<Path>) -> crate::result::Result<Self> {
+        let dict_dir = open_jtalk_dict_dir
+            .as_ref()
+            .to_str()
+            .unwrap_or_else(|| todo!()) // FIXME: `camino::Utf8Path`を要求するようにする
+            .to_owned();
+        let dict_dir = Arc::new(dict_dir);
+
+        crate::task::asyncify(move || {
+            let mut resources = Resources {
                 mecab: ManagedResource::initialize(),
                 njd: ManagedResource::initialize(),
                 jpcommon: ManagedResource::initialize(),
-            })
-            .into(),
-            dict_dir: None,
-        }
-    }
+            };
 
-    pub async fn new(open_jtalk_dict_dir: impl AsRef<Path>) -> crate::result::Result<Self> {
-        let open_jtalk_dict_dir = open_jtalk_dict_dir.as_ref().to_owned();
-
-        crate::task::asyncify(move || {
-            let mut s = Self::new_without_dic();
-            s.load(open_jtalk_dict_dir).map_err(|()| {
+            let result = resources.mecab.load(&*dict_dir);
+            if !result {
                 // FIXME: 「システム辞書を読もうとしたけど読めなかった」というエラーをちゃんと用意する
-                ErrorRepr::NotLoadedOpenjtalkDict
-            })?;
-            Ok(s)
+                return Err(ErrorRepr::NotLoadedOpenjtalkDict.into());
+            }
+
+            Ok(Self {
+                resources: Mutex::new(resources).into(),
+                dict_dir,
+            })
         })
         .await
     }
 
-    // 先に`load`を呼ぶ必要がある。
     /// ユーザー辞書を設定する。
     ///
     /// この関数を呼び出した後にユーザー辞書を変更した場合は、再度この関数を呼ぶ必要がある。
     pub async fn use_user_dict(&self, user_dict: &UserDict) -> crate::result::Result<()> {
-        let dict_dir = self
-            .dict_dir
-            .as_ref()
-            .and_then(|dict_dir| dict_dir.to_str())
-            .ok_or(ErrorRepr::NotLoadedOpenjtalkDict)?
-            .to_owned();
-
         let resources = self.resources.clone();
+        let dict_dir = self.dict_dir.clone();
 
         let words = user_dict.to_mecab_format();
 
@@ -108,7 +101,7 @@ impl OpenJtalk {
 
             let Resources { mecab, .. } = &mut *resources.lock().unwrap();
 
-            Ok(mecab.load_with_userdic(dict_dir.as_ref(), Some(Path::new(&temp_dict_path))))
+            Ok(mecab.load_with_userdic((*dict_dir).as_ref(), Some(Path::new(&temp_dict_path))))
         })
         .await?;
 
@@ -167,26 +160,6 @@ impl OpenJtalk {
             })
         }
     }
-
-    fn load(&mut self, open_jtalk_dict_dir: impl AsRef<Path>) -> std::result::Result<(), ()> {
-        let result = self
-            .resources
-            .lock()
-            .unwrap()
-            .mecab
-            .load(open_jtalk_dict_dir.as_ref());
-        if result {
-            self.dict_dir = Some(open_jtalk_dict_dir.as_ref().into());
-            Ok(())
-        } else {
-            self.dict_dir = None;
-            Err(())
-        }
-    }
-
-    pub fn dict_loaded(&self) -> bool {
-        self.dict_dir.is_some()
-    }
 }
 
 #[cfg(test)]
diff --git a/crates/voicevox_core/src/synthesizer.rs b/crates/voicevox_core/src/synthesizer.rs
index 383e8f6c0..556877e08 100644
--- a/crates/voicevox_core/src/synthesizer.rs
+++ b/crates/voicevox_core/src/synthesizer.rs
@@ -3,7 +3,6 @@ use std::{
     sync::Arc,
 };
 
-use easy_ext::ext;
 use enum_map::enum_map;
 
 use crate::{
@@ -95,11 +94,11 @@ pub(crate) type InferenceRuntimeImpl = Onnxruntime;
 
 /// 音声シンセサイザ。
 #[derive(Clone)]
-pub struct Synthesizer(Arc<blocking::Synthesizer>);
+pub struct Synthesizer<O>(Arc<blocking::Synthesizer<O>>);
 
 // FIXME: docを書く
-impl Synthesizer {
-    pub fn new(open_jtalk: Arc<OpenJtalk>, options: &InitializeOptions) -> Result<Self> {
+impl<O: Send + Sync + 'static> Synthesizer<O> {
+    pub fn new(open_jtalk: O, options: &InitializeOptions) -> Result<Self> {
         blocking::Synthesizer::new(open_jtalk, options)
             .map(Into::into)
             .map(Self)
@@ -155,17 +154,6 @@ impl Synthesizer {
             .await
     }
 
-    pub async fn create_accent_phrases(
-        &self,
-        text: &str,
-        style_id: StyleId,
-    ) -> Result<Vec<AccentPhraseModel>> {
-        let blocking = self.0.clone();
-        let text = text.to_owned();
-
-        crate::task::asyncify(move || blocking.create_accent_phrases(&text, style_id)).await
-    }
-
     pub async fn replace_mora_data(
         &self,
         accent_phrases: &[AccentPhraseModel],
@@ -211,13 +199,6 @@ impl Synthesizer {
         crate::task::asyncify(move || blocking.audio_query_from_kana(&kana, style_id)).await
     }
 
-    pub async fn audio_query(&self, text: &str, style_id: StyleId) -> Result<AudioQueryModel> {
-        let blocking = self.0.clone();
-        let text = text.to_owned();
-
-        crate::task::asyncify(move || blocking.audio_query(&text, style_id)).await
-    }
-
     pub async fn tts_from_kana(
         &self,
         kana: &str,
@@ -230,6 +211,26 @@ impl Synthesizer {
 
         crate::task::asyncify(move || blocking.tts_from_kana(&kana, style_id, &options)).await
     }
+}
+
+impl Synthesizer<OpenJtalk> {
+    pub async fn create_accent_phrases(
+        &self,
+        text: &str,
+        style_id: StyleId,
+    ) -> Result<Vec<AccentPhraseModel>> {
+        let blocking = self.0.clone();
+        let text = text.to_owned();
+
+        crate::task::asyncify(move || blocking.create_accent_phrases(&text, style_id)).await
+    }
+
+    pub async fn audio_query(&self, text: &str, style_id: StyleId) -> Result<AudioQueryModel> {
+        let blocking = self.0.clone();
+        let text = text.to_owned();
+
+        crate::task::asyncify(move || blocking.audio_query(&text, style_id)).await
+    }
 
     pub async fn tts(
         &self,
@@ -245,9 +246,12 @@ impl Synthesizer {
     }
 }
 
+// FIXME: コードのdiffを抑えるため`impl<O> blocking::Synthesizer<O>`と
+// `impl blocking::Synthesizer<OpenJtalk>`がそれぞれ3つ誕生しているので、一つずつにまとめる
+
 // FIXME: ここのdocのコードブロックはasync版のものなので、↑の方に移した上で、(ブロッキング版を
 // public APIにするならの話ではあるが)ブロッキング版はブロッキング版でコード例を用意する
-impl blocking::Synthesizer {
+impl<O> blocking::Synthesizer<O> {
     /// `Synthesizer`をコンストラクトする。
     ///
     /// # Example
@@ -275,7 +279,7 @@ impl blocking::Synthesizer {
     /// # Ok(())
     /// # }
     /// ```
-    fn new(open_jtalk: Arc<OpenJtalk>, options: &InitializeOptions) -> Result<Self> {
+    fn new(open_jtalk: O, options: &InitializeOptions) -> Result<Self> {
         #[cfg(windows)]
         list_windows_video_cards();
 
@@ -599,7 +603,9 @@ impl blocking::Synthesizer {
     ) -> Result<Vec<AccentPhraseModel>> {
         self.replace_mora_data(&parse_kana(kana)?, style_id)
     }
+}
 
+impl blocking::Synthesizer<OpenJtalk> {
     /// 日本語のテキストからAccentPhrase (アクセント句)の配列を生成する。
     ///
     /// # Example
@@ -628,10 +634,6 @@ impl blocking::Synthesizer {
         text: &str,
         style_id: StyleId,
     ) -> Result<Vec<AccentPhraseModel>> {
-        if !self.open_jtalk.dict_loaded() {
-            return Err(ErrorRepr::NotLoadedOpenjtalkDict.into());
-        }
-
         if text.is_empty() {
             return Ok(Vec::new());
         }
@@ -703,7 +705,9 @@ impl blocking::Synthesizer {
 
         self.replace_mora_data(&accent_phrases, style_id)
     }
+}
 
+impl<O> blocking::Synthesizer<O> {
     /// AccentPhraseの配列の音高・音素長を、特定の声で生成しなおす。
     fn replace_mora_data(
         &self,
@@ -939,7 +943,9 @@ impl blocking::Synthesizer {
         let accent_phrases = self.create_accent_phrases_from_kana(kana, style_id)?;
         Ok(AudioQueryModel::from_accent_phrases(accent_phrases).with_kana(Some(kana.to_owned())))
     }
+}
 
+impl blocking::Synthesizer<OpenJtalk> {
     /// 日本語のテキストから[AudioQuery]を生成する。
     ///
     /// # Examples
@@ -969,7 +975,9 @@ impl blocking::Synthesizer {
         let accent_phrases = self.create_accent_phrases(text, style_id)?;
         Ok(AudioQueryModel::from_accent_phrases(accent_phrases))
     }
+}
 
+impl<O> blocking::Synthesizer<O> {
     /// AquesTalk風記法から音声合成を行う。
     fn tts_from_kana(
         &self,
@@ -980,7 +988,9 @@ impl blocking::Synthesizer {
         let audio_query = &self.audio_query_from_kana(kana, style_id)?;
         self.synthesis(audio_query, style_id, &SynthesisOptions::from(options))
     }
+}
 
+impl blocking::Synthesizer<OpenJtalk> {
     /// 日本語のテキストから音声合成を行う。
     fn tts(&self, text: &str, style_id: StyleId, options: &TtsOptions) -> Result<Vec<u8>> {
         let audio_query = &self.audio_query(text, style_id)?;
@@ -988,7 +998,48 @@ impl blocking::Synthesizer {
     }
 }
 
-impl PerformInference for Synthesizer {
+pub trait PerformInference {
+    /// `predict_duration`を実行する。
+    ///
+    /// # Performance
+    ///
+    /// CPU-boundな操作であるため、非同期ランタイム上では直接実行されるべきではない。
+    fn predict_duration(&self, phoneme_vector: &[i64], style_id: StyleId) -> Result<Vec<f32>>;
+
+    /// `predict_intonation`を実行する。
+    ///
+    /// # Performance
+    ///
+    /// CPU-boundな操作であるため、非同期ランタイム上では直接実行されるべきではない。
+    #[allow(clippy::too_many_arguments)]
+    fn predict_intonation(
+        &self,
+        length: usize,
+        vowel_phoneme_vector: &[i64],
+        consonant_phoneme_vector: &[i64],
+        start_accent_vector: &[i64],
+        end_accent_vector: &[i64],
+        start_accent_phrase_vector: &[i64],
+        end_accent_phrase_vector: &[i64],
+        style_id: StyleId,
+    ) -> Result<Vec<f32>>;
+
+    /// `decode`を実行する。
+    ///
+    /// # Performance
+    ///
+    /// CPU/GPU-boundな操作であるため、非同期ランタイム上では直接実行されるべきではない。
+    fn decode(
+        &self,
+        length: usize,
+        phoneme_size: usize,
+        f0: &[f32],
+        phoneme_vector: &[f32],
+        style_id: StyleId,
+    ) -> Result<Vec<f32>>;
+}
+
+impl<O> PerformInference for Synthesizer<O> {
     fn predict_duration(&self, phoneme_vector: &[i64], style_id: StyleId) -> Result<Vec<f32>> {
         self.0.predict_duration(phoneme_vector, style_id)
     }
@@ -1029,14 +1080,8 @@ impl PerformInference for Synthesizer {
     }
 }
 
-#[ext(PerformInference)]
-impl blocking::Synthesizer {
-    /// `predict_duration`を実行する。
-    ///
-    /// # Performance
-    ///
-    /// CPU-boundな操作であるため、非同期ランタイム上では直接実行されるべきではない。
-    pub fn predict_duration(&self, phoneme_vector: &[i64], style_id: StyleId) -> Result<Vec<f32>> {
+impl<O> PerformInference for blocking::Synthesizer<O> {
+    fn predict_duration(&self, phoneme_vector: &[i64], style_id: StyleId) -> Result<Vec<f32>> {
         // FIXME: `Status::ids_for`があるため、ここは不要なはず
         if !self.status.validate_speaker_id(style_id) {
             return Err(ErrorRepr::StyleNotFound { style_id }.into());
@@ -1066,13 +1111,7 @@ impl blocking::Synthesizer {
         const PHONEME_LENGTH_MINIMAL: f32 = 0.01;
     }
 
-    /// `predict_intonation`を実行する。
-    ///
-    /// # Performance
-    ///
-    /// CPU-boundな操作であるため、非同期ランタイム上では直接実行されるべきではない。
-    #[allow(clippy::too_many_arguments)]
-    pub fn predict_intonation(
+    fn predict_intonation(
         &self,
         length: usize,
         vowel_phoneme_vector: &[i64],
@@ -1107,12 +1146,7 @@ impl blocking::Synthesizer {
         Ok(output.into_raw_vec())
     }
 
-    /// `decode`を実行する。
-    ///
-    /// # Performance
-    ///
-    /// CPU/GPU-boundな操作であるため、非同期ランタイム上では直接実行されるべきではない。
-    pub fn decode(
+    fn decode(
         &self,
         length: usize,
         phoneme_size: usize,
@@ -1360,18 +1394,13 @@ impl AudioQueryModel {
 }
 
 mod blocking {
-    use std::sync::Arc;
-
-    use crate::{
-        engine::OpenJtalk,
-        infer::{domain::InferenceDomainImpl, status::Status},
-    };
+    use crate::infer::{domain::InferenceDomainImpl, status::Status};
 
     use super::InferenceRuntimeImpl;
 
-    pub(super) struct Synthesizer {
+    pub(super) struct Synthesizer<O> {
         pub(super) status: Status<InferenceRuntimeImpl, InferenceDomainImpl>,
-        pub(super) open_jtalk: Arc<OpenJtalk>,
+        pub(super) open_jtalk: O,
         pub(super) use_gpu: bool,
     }
 }
@@ -1388,7 +1417,7 @@ mod tests {
     #[tokio::test]
     async fn load_model_works(#[case] expected_result_at_initialized: Result<()>) {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new_without_dic()),
+            (),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1411,7 +1440,7 @@ mod tests {
     #[tokio::test]
     async fn is_use_gpu_works() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new_without_dic()),
+            (),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1427,7 +1456,7 @@ mod tests {
     async fn is_loaded_model_by_style_id_works(#[case] style_id: u32, #[case] expected: bool) {
         let style_id = StyleId::new(style_id);
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new_without_dic()),
+            (),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1455,7 +1484,7 @@ mod tests {
     #[tokio::test]
     async fn predict_duration_works() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new_without_dic()),
+            (),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1484,7 +1513,7 @@ mod tests {
     #[tokio::test]
     async fn predict_intonation_works() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new_without_dic()),
+            (),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1523,7 +1552,7 @@ mod tests {
     #[tokio::test]
     async fn decode_works() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new_without_dic()),
+            (),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1613,7 +1642,7 @@ mod tests {
         #[case] expected_kana_text: &str,
     ) {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap()),
+            OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap(),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1681,7 +1710,7 @@ mod tests {
         #[case] expected_text_consonant_vowel_data: &TextConsonantVowelData,
     ) {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap()),
+            OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap(),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1746,7 +1775,7 @@ mod tests {
     #[tokio::test]
     async fn create_accent_phrases_works_for_japanese_commas_and_periods() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap()),
+            OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap(),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1805,7 +1834,7 @@ mod tests {
     #[tokio::test]
     async fn mora_length_works() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap()),
+            OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap(),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1841,7 +1870,7 @@ mod tests {
     #[tokio::test]
     async fn mora_pitch_works() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap()),
+            OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap(),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
@@ -1873,7 +1902,7 @@ mod tests {
     #[tokio::test]
     async fn mora_data_works() {
         let syntesizer = Synthesizer::new(
-            Arc::new(OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap()),
+            OpenJtalk::new(OPEN_JTALK_DIC_DIR).await.unwrap(),
             &InitializeOptions {
                 acceleration_mode: AccelerationMode::Cpu,
                 ..Default::default()
diff --git a/crates/voicevox_core_c_api/src/c_impls.rs b/crates/voicevox_core_c_api/src/c_impls.rs
index 74e783fdc..aa1421049 100644
--- a/crates/voicevox_core_c_api/src/c_impls.rs
+++ b/crates/voicevox_core_c_api/src/c_impls.rs
@@ -1,4 +1,4 @@
-use std::{ffi::CString, path::Path, sync::Arc};
+use std::{ffi::CString, path::Path};
 
 use voicevox_core::{InitializeOptions, OpenJtalk, Result, Synthesizer, VoiceModel, VoiceModelId};
 
@@ -7,7 +7,7 @@ use crate::{CApiResult, OpenJtalkRc, VoicevoxSynthesizer, VoicevoxVoiceModel};
 impl OpenJtalkRc {
     pub(crate) async fn new(open_jtalk_dic_dir: impl AsRef<Path>) -> Result<Self> {
         Ok(Self {
-            open_jtalk: Arc::new(OpenJtalk::new(open_jtalk_dic_dir).await?),
+            open_jtalk: OpenJtalk::new(open_jtalk_dic_dir).await?,
         })
     }
 }
diff --git a/crates/voicevox_core_c_api/src/compatible_engine.rs b/crates/voicevox_core_c_api/src/compatible_engine.rs
index 8f2d7c5cc..cf19d9d3b 100644
--- a/crates/voicevox_core_c_api/src/compatible_engine.rs
+++ b/crates/voicevox_core_c_api/src/compatible_engine.rs
@@ -1,9 +1,9 @@
-use std::{collections::BTreeMap, sync::Arc};
+use std::collections::BTreeMap;
 
 use super::*;
 use libc::c_int;
 
-use voicevox_core::{OpenJtalk, StyleId, VoiceModel, __internal::interop::PerformInference as _};
+use voicevox_core::{StyleId, VoiceModel, __internal::interop::PerformInference as _};
 
 macro_rules! ensure_initialized {
     ($synthesizer:expr $(,)?) => {
@@ -88,10 +88,10 @@ fn voice_model_set() -> &'static VoiceModelSet {
     &VOICE_MODEL_SET
 }
 
-static SYNTHESIZER: Lazy<Mutex<Option<voicevox_core::Synthesizer>>> =
+static SYNTHESIZER: Lazy<Mutex<Option<voicevox_core::Synthesizer<()>>>> =
     Lazy::new(|| Mutex::new(None));
 
-fn lock_synthesizer() -> MutexGuard<'static, Option<voicevox_core::Synthesizer>> {
+fn lock_synthesizer() -> MutexGuard<'static, Option<voicevox_core::Synthesizer<()>>> {
     SYNTHESIZER.lock().unwrap()
 }
 
@@ -108,7 +108,7 @@ pub extern "C" fn initialize(use_gpu: bool, cpu_num_threads: c_int, load_all_mod
     // で行っているという構造になってしまっているので、外すとロガーの初期化が遅れてしまでう
     let result = RUNTIME.block_on(async {
         let synthesizer = voicevox_core::Synthesizer::new(
-            Arc::new(OpenJtalk::new_without_dic()),
+            (),
             &voicevox_core::InitializeOptions {
                 acceleration_mode: if use_gpu {
                     voicevox_core::AccelerationMode::Gpu
diff --git a/crates/voicevox_core_c_api/src/lib.rs b/crates/voicevox_core_c_api/src/lib.rs
index ac72b6359..b9e50ba5d 100644
--- a/crates/voicevox_core_c_api/src/lib.rs
+++ b/crates/voicevox_core_c_api/src/lib.rs
@@ -104,7 +104,7 @@ static RUNTIME: Lazy<Runtime> = Lazy::new(|| {
 /// ```
 /// }
 pub struct OpenJtalkRc {
-    open_jtalk: Arc<OpenJtalk>,
+    open_jtalk: OpenJtalk,
 }
 
 /// ::OpenJtalkRc を<b>構築</b>(_construct_)する。
@@ -317,7 +317,7 @@ pub extern "C" fn voicevox_voice_model_delete(model: Box<VoicevoxVoiceModel>) {
 /// <b>構築</b>(_construction_)は ::voicevox_synthesizer_new で行い、<b>破棄</b>(_destruction_)は ::voicevox_synthesizer_delete で行う。
 #[derive(Getters)]
 pub struct VoicevoxSynthesizer {
-    synthesizer: Synthesizer,
+    synthesizer: Synthesizer<OpenJtalk>,
 }
 
 /// ::VoicevoxSynthesizer を<b>構築</b>(_construct_)する。
diff --git a/crates/voicevox_core_java_api/src/open_jtalk.rs b/crates/voicevox_core_java_api/src/open_jtalk.rs
index bbb03608c..422dd0fa8 100644
--- a/crates/voicevox_core_java_api/src/open_jtalk.rs
+++ b/crates/voicevox_core_java_api/src/open_jtalk.rs
@@ -17,7 +17,7 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_OpenJtalk_rsNew<'local>
         let open_jtalk_dict_dir = &*Cow::from(&open_jtalk_dict_dir);
 
         let internal = RUNTIME.block_on(voicevox_core::OpenJtalk::new(open_jtalk_dict_dir))?;
-        env.set_rust_field(&this, "handle", Arc::new(internal))?;
+        env.set_rust_field(&this, "handle", internal)?;
 
         Ok(())
     })
@@ -31,7 +31,7 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_OpenJtalk_rsUseUserDict
 ) {
     throw_if_err(env, (), |env| {
         let internal = env
-            .get_rust_field::<_, _, Arc<voicevox_core::OpenJtalk>>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::OpenJtalk>(&this, "handle")?
             .clone();
 
         let user_dict = env
diff --git a/crates/voicevox_core_java_api/src/synthesizer.rs b/crates/voicevox_core_java_api/src/synthesizer.rs
index 5e2cbb8d2..8828c6e78 100644
--- a/crates/voicevox_core_java_api/src/synthesizer.rs
+++ b/crates/voicevox_core_java_api/src/synthesizer.rs
@@ -50,7 +50,7 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsNew<'loca
         options.cpu_num_threads = cpu_num_threads.i().expect("cpuNumThreads is not integer") as u16;
 
         let open_jtalk = env
-            .get_rust_field::<_, _, Arc<voicevox_core::OpenJtalk>>(&open_jtalk, "handle")?
+            .get_rust_field::<_, _, voicevox_core::OpenJtalk>(&open_jtalk, "handle")?
             .clone();
         let internal = voicevox_core::Synthesizer::new(open_jtalk, Box::leak(Box::new(options)))?;
         env.set_rust_field(&this, "handle", internal)?;
@@ -64,7 +64,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsIsGpuMode
 ) -> jboolean {
     throw_if_err(env, false, |env| {
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         Ok(internal.is_gpu_mode())
@@ -78,7 +80,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsGetMetasJ
 ) -> jobject {
     throw_if_err(env, std::ptr::null_mut(), |env| {
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let metas_json = serde_json::to_string(&internal.metas()).expect("should not fail");
@@ -100,7 +104,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsLoadVoice
             .get_rust_field::<_, _, Arc<voicevox_core::VoiceModel>>(&model, "handle")?
             .clone();
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
         RUNTIME.block_on(internal.load_voice_model(&model))?;
         Ok(())
@@ -117,7 +123,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsUnloadVoi
         let model_id: String = env.get_string(&model_id)?.into();
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         internal.unload_voice_model(&voicevox_core::VoiceModelId::new(model_id))?;
@@ -138,7 +146,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsIsLoadedV
         let model_id: String = env.get_string(&model_id)?.into();
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let is_loaded = internal.is_loaded_voice_model(&voicevox_core::VoiceModelId::new(model_id));
@@ -162,7 +172,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsAudioQuer
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let audio_query = RUNTIME.block_on(
@@ -189,7 +201,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsAudioQuer
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let audio_query =
@@ -217,7 +231,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsAccentPhr
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let accent_phrases = RUNTIME.block_on(
@@ -244,7 +260,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsAccentPhr
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let accent_phrases = RUNTIME.block_on(
@@ -273,7 +291,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsReplaceMo
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let replaced_accent_phrases = RUNTIME.block_on(
@@ -303,7 +323,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsReplacePh
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let replaced_accent_phrases = {
@@ -334,7 +356,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsReplaceMo
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let replaced_accent_phrases = RUNTIME.block_on(
@@ -363,7 +387,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsSynthesis
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let wave = {
@@ -397,7 +423,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsTtsFromKa
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let wave = {
@@ -431,7 +459,9 @@ unsafe extern "system" fn Java_jp_hiroshiba_voicevoxcore_Synthesizer_rsTts<'loca
         let style_id = style_id as u32;
 
         let internal = env
-            .get_rust_field::<_, _, voicevox_core::Synthesizer>(&this, "handle")?
+            .get_rust_field::<_, _, voicevox_core::Synthesizer<voicevox_core::OpenJtalk>>(
+                &this, "handle",
+            )?
             .clone();
 
         let wave = {
diff --git a/crates/voicevox_core_python_api/src/lib.rs b/crates/voicevox_core_python_api/src/lib.rs
index 86a64e394..956c924b2 100644
--- a/crates/voicevox_core_python_api/src/lib.rs
+++ b/crates/voicevox_core_python_api/src/lib.rs
@@ -114,7 +114,7 @@ impl VoiceModel {
 #[pyclass]
 #[derive(Clone)]
 struct OpenJtalk {
-    open_jtalk: Arc<voicevox_core::OpenJtalk>,
+    open_jtalk: voicevox_core::OpenJtalk,
 }
 
 #[pymethods]
@@ -127,7 +127,7 @@ impl OpenJtalk {
     ) -> PyResult<&PyAny> {
         pyo3_asyncio::tokio::future_into_py(py, async move {
             let open_jtalk = voicevox_core::OpenJtalk::new(open_jtalk_dict_dir).await;
-            let open_jtalk = Python::with_gil(|py| open_jtalk.into_py_result(py))?.into();
+            let open_jtalk = Python::with_gil(|py| open_jtalk.into_py_result(py))?;
             Ok(Self { open_jtalk })
         })
     }
@@ -144,7 +144,7 @@ impl OpenJtalk {
 
 #[pyclass]
 struct Synthesizer {
-    synthesizer: Closable<voicevox_core::Synthesizer, Self>,
+    synthesizer: Closable<voicevox_core::Synthesizer<voicevox_core::OpenJtalk>, Self>,
 }
 
 #[pymethods]