diff --git a/Cargo.toml b/Cargo.toml
index eea4c66..97ccde7 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -3,3 +3,4 @@ members = [
     "vosk",
     "vosk-sys",
 ]
+resolver = "2"
diff --git a/vosk-sys/src/lib.rs b/vosk-sys/src/lib.rs
index 2a5b4e3..b870bbc 100644
--- a/vosk-sys/src/lib.rs
+++ b/vosk-sys/src/lib.rs
@@ -304,7 +304,7 @@ extern "C" {
     #[doc = " Creates the batch recognizer object"]
     #[doc = ""]
     #[doc = "  @returns model object or NULL if problem occured"]
-    pub fn vosk_batch_model_new() -> *mut VoskBatchModel;
+    pub fn vosk_batch_model_new(model_path: *const ::std::os::raw::c_char) -> *mut VoskBatchModel;
 
     #[doc = " Releases batch model object"]
     pub fn vosk_batch_model_free(model: *mut VoskBatchModel);
diff --git a/vosk/Cargo.toml b/vosk/Cargo.toml
index 8117b60..19c4334 100644
--- a/vosk/Cargo.toml
+++ b/vosk/Cargo.toml
@@ -19,3 +19,6 @@ serde = { version = "1.0", features = ["derive"] }
 cpal = "0.14"
 dasp = "0.11"
 hound = "3.5"
+
+[features]
+cuda = []
diff --git a/vosk/src/lib.rs b/vosk/src/lib.rs
index af10495..9b765d3 100644
--- a/vosk/src/lib.rs
+++ b/vosk/src/lib.rs
@@ -15,3 +15,17 @@ mod recognition;
 pub use log::*;
 pub use models::*;
 pub use recognition::*;
+
+/// Init, automatically select a CUDA device and allow multithreading.
+/// Must be called once from the main thread.
+#[cfg(feature = "cuda")]
+pub fn gpu_init() {
+    unsafe { vosk_sys::vosk_gpu_init() }
+}
+
+/// Init CUDA device in a multi-threaded environment.
+/// Must be called for each thread.
+#[cfg(feature = "cuda")]
+pub fn gpu_thread_init() {
+    unsafe { vosk_sys::vosk_gpu_thread_init() }
+}
diff --git a/vosk/src/models.rs b/vosk/src/models.rs
index 8cfa04f..b5a5046 100644
--- a/vosk/src/models.rs
+++ b/vosk/src/models.rs
@@ -76,3 +76,41 @@ impl Drop for SpeakerModel {
 
 unsafe impl Send for SpeakerModel {}
 unsafe impl Sync for SpeakerModel {}
+
+#[cfg(feature = "cuda")]
+pub mod batch_model {
+    use std::{ffi::CString, ptr::NonNull};
+    use vosk_sys::*;
+    /// The same as [`Model`], but uses a CUDA enabled Nvidia GPU and dynamic batching to enable higher throughput.
+
+    pub struct BatchModel(pub(crate) NonNull<VoskBatchModel>);
+
+    impl BatchModel {
+        /// Loads model data from the file and returns the model object, or [`None`]
+        /// if a problem occured.
+        ///
+        /// * `model_path` - the path to the model directory.
+        #[must_use]
+        pub fn new(model_path: impl Into<String>) -> Option<Self> {
+            let model_path_c = CString::new(model_path.into()).ok()?;
+            let model_ptr = unsafe { vosk_batch_model_new(model_path_c.as_ptr()) };
+
+            Some(Self(NonNull::new(model_ptr)?))
+        }
+
+        /// Waits for inferencing to finish
+        pub fn wait(&self) {
+            unsafe { vosk_batch_model_wait(self.0.as_ptr()) };
+        }
+    }
+
+    impl Drop for BatchModel {
+        fn drop(&mut self) {
+            unsafe { vosk_batch_model_free(self.0.as_ptr()) }
+        }
+    }
+
+    unsafe impl Send for BatchModel {}
+
+    unsafe impl Sync for BatchModel {}
+}
diff --git a/vosk/src/recognition/mod.rs b/vosk/src/recognition/mod.rs
index 9e1d4e1..d34f14d 100644
--- a/vosk/src/recognition/mod.rs
+++ b/vosk/src/recognition/mod.rs
@@ -1,4 +1,5 @@
 use crate::{Model, SpeakerModel};
+
 use serde::Deserialize;
 use std::{
     ffi::{CStr, CString},
@@ -189,6 +190,11 @@ impl Recognizer {
         unsafe { vosk_recognizer_set_partial_words(self.0.as_ptr(), i32::from(enable)) }
     }
 
+    /// Enables or disables Natural Language Semantics Markup Language (NLSML) in the output
+    pub fn set_nlsml(&mut self, enable: bool) {
+        unsafe { vosk_recognizer_set_nlsml(self.0.as_ptr(), i32::from(enable)) }
+    }
+
     /// Accept and process new chunk of voice data.
     ///
     /// * `data` - Audio data in PCM 16-bit mono format.
@@ -277,3 +283,86 @@ impl Drop for Recognizer {
         unsafe { vosk_recognizer_free(self.0.as_ptr()) }
     }
 }
+
+#[cfg(feature = "cuda")]
+pub mod batch_recognizer {
+    use crate::batch_model::BatchModel;
+    use vosk_sys::*;
+
+    use std::{ffi::CStr, ptr::NonNull};
+
+    pub use crate::recognition::results::*;
+
+    /// The main object which processes data using GPU inferencing.
+    /// Takes audio as input and returns decoded information as words, confidences, times, and other metadata.
+
+    pub struct BatchRecognizer(std::ptr::NonNull<VoskBatchRecognizer>);
+
+    impl BatchRecognizer {
+        /// Creates the recognizer object. Returns [`None`] if a problem occured.
+        ///
+        /// The recognizers process the speech and return text using shared model data.
+        ///
+        /// * `model` - [`BatchModel`] containing static data for recognizer. Model can be shared
+        /// across recognizers, even running in different threads.
+        ///
+        /// * `sample_rate` - The sample rate of the audio you going to feed into the recognizer.
+        /// Make sure this rate matches the audio content, it is a common issue causing accuracy problems.
+        ///
+        /// [`BatchModel`]: crate::BatchModel
+        #[must_use]
+        pub fn new(model: &BatchModel, sample_rate: f32) -> Option<Self> {
+            let recognizer_ptr =
+                unsafe { vosk_batch_recognizer_new(model.0.as_ptr(), sample_rate) };
+            Some(Self(NonNull::new(recognizer_ptr)?))
+        }
+
+        /// Enables or disables Natural Language Semantics Markup Language (NLSML) in the output
+        pub fn set_nlsml(&mut self, enable: bool) {
+            unsafe { vosk_batch_recognizer_set_nlsml(self.0.as_ptr(), i32::from(enable)) }
+        }
+
+        /// Accept and process new chunk of voice data.
+        ///
+        /// * `data` - Audio data in PCM 16-bit mono format as an array of i8.
+        pub fn accept_waveform(&mut self, data: &[i8]) {
+            unsafe {
+                vosk_batch_recognizer_accept_waveform(
+                    self.0.as_ptr(),
+                    data.as_ptr(),
+                    data.len() as i32,
+                )
+            };
+        }
+
+        /// Closes the stream to the model
+        pub fn finish_stream(&mut self) {
+            unsafe { vosk_batch_recognizer_finish_stream(self.0.as_ptr()) };
+        }
+
+        /// Gets the front of the result queue
+        pub fn front_result(&mut self) -> Result<Word, serde_json::Error> {
+            serde_json::from_str(
+                unsafe { CStr::from_ptr(vosk_batch_recognizer_front_result(self.0.as_ptr())) }
+                    .to_str()
+                    .unwrap(),
+            )
+        }
+
+        /// Removes the front of the result queue
+        pub fn pop(&mut self) {
+            unsafe { vosk_batch_recognizer_pop(self.0.as_ptr()) }
+        }
+
+        /// Gets the number of chunks that have yet to be processed
+        pub fn get_pending_chunks(&mut self) -> i32 {
+            (unsafe { vosk_batch_recognizer_get_pending_chunks(self.0.as_ptr()) }) as i32
+        }
+    }
+
+    impl Drop for BatchRecognizer {
+        fn drop(&mut self) {
+            unsafe { vosk_batch_recognizer_free(self.0.as_ptr()) }
+        }
+    }
+}