diff --git a/soundtrigger/2.4/ISoundTriggerHw.hal b/soundtrigger/2.4/ISoundTriggerHw.hal
new file mode 100644
index 0000000..fd39303
--- /dev/null
+++ b/soundtrigger/2.4/ISoundTriggerHw.hal
@@ -0,0 +1,190 @@
+/*
+ * Copyright 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package android.hardware.soundtrigger@2.4;
+
+import @2.0::SoundModelHandle;
+import @2.1::ISoundTriggerHw.SoundModel;
+import @2.1::ISoundTriggerHw.PhraseSoundModel;
+import @2.3::ISoundTriggerHw;
+import @2.3::RecognitionConfig;
+import ISoundTriggerHwCallback;
+import ISoundTriggerHwGlobalCallback;
+
+/**
+ * SoundTrigger HAL interface. Used for hardware recognition of hotwords
+ * and other sounds.
+ *
+ * Important notes about the threading model:
+ * ==========================================
+ * Both this interface and the corresponding callback interface use a synchronized calling
+ * convention. This model comes with some advantages, but also with some risks of deadlocks if the
+ * implementation does not handle this correctly. Please consider the following:
+ * - After stopRecognition() returns no more recognition events for that model may be sent. This
+ *   implies that any queues holding such events must be flushed before the call returns and that
+ *   may imply that callback from the HAL to the client are done while stopRecognition() is blocked.
+ *   This is OK, and supported by the framework.
+ * - Similarly, the same relationship applies between unloadModel() and subsequent callbacks to
+ *   modelUnloaded().
+ * - Other than these two cases, calls into the HAL *MAY NOT* block on callbacks from the HAL, or
+ *   else deadlock conditions may result, which may be handled by rebooting of the HAL process and
+ *   cause service outages.
+ *
+ * Similarly, it is expected that a single call to startRecognition() generates at most one event
+ * (the model automatically becomes stopped when the event occurs, until explicitly started again)
+ * and that after a modelUnloaded() event no more events would be sent regarding the model.
+ * Note that a getModelState() call may generate a recognition event, but this event DOES NOT modify
+ * the model state - the model remains started.
+ *
+ * The HAL is expected to correctly handle a stopRecognition() call even after it sent an event
+ * indicating that recognition is stopped and an unloadModel() call even after it sent an event
+ * indicating that it has been unloaded. This is required in order to prevent race conditions
+ * between these calls. This also implies that model handles should generally not be reused until
+ * explicitly unloaded. To avoid the rare possibility of running out of handles, the framework will
+ * call unloadModel() on models that have been preemptively unloaded by the HAL.
+ *
+ * Due to the asynchronous nature of recognition events and preemptive model unloading, the HAL must
+ * correctly handle requests that would have been valid before an event has been delivered, but
+ * became moot as result of the event. Namely:
+ * - stopRecognition() may be called on a model that has already delivered an event and became
+ *   inactive as a result. The HAL must return a successful return code in this case.
+ * - Furthermore, if a model is preemptively unloaded after it triggers (typically, this would
+ *   happen when it is first aborted and immediately preemptively unloaded), stopRecognition() may
+ *   be called on it. The HAL must return a successful return code in this case.
+ * - startRecognition() may be called on a model that has been preemptively unloaded. In this case,
+ *   the HAL must return -EBUSY to indicate that the operation is temporarily unsuccessful.
+ * - unloadSoundModel() may be called on a model that has been preemptively unloaded. The HAL must
+ *   return a successful return code in this case.
+ *
+ * Important notes about resource constraints and concurrency
+ * =========================================================
+ * Up until this version, the framework would enforce concurrency constraints expressed by the
+ * Properties presented by the soundtrigger instance. These include constraints on the maximum
+ * amount of models that can be loaded at the same time and on running recognition while capturing
+ * from the microphone.
+ * This version changes the approach for how these constraints are modeled, both offering the HAL
+ * implementation more flexibility and simplifying the framework's job in enforcing these
+ * limitations. Note that there is no change for how the framework behaves with earlier versions,
+ * everything described below only applies to this version and onward.
+ * The way this is achieved is as following:
+ * - The framework will no longer enforce constraints on concurrent loading of models, as expressed
+ *   in the Properties.maxSoundModels field (this property is merely a hint at this point and may be
+ *   deprecated in the future.
+ * - The framework will no longer enforce constraints on concurrency of audio recording and
+ *   soundtrigger operation, as expressed in the Properties.concurrentCapture field (this property
+ *   is merely a hint at this point and may be deprecated in the future).
+ * - The framework will no longer enforce constraints on concurrent loading of models, as expressed
+ *   in the Properties (these properties are merely hints at this point and may be deprecated in the
+ *   future.
+ * - The HAL implementation is free to reject starting of any model at any time by having the
+ *   respective start*() method return -EBUSY.
+ * - The HAL implementation is free to reject loading of any model at any time by having the
+ *   respective load*() method return -EBUSY.
+ * - The HAL implementation is free to preemptively stop a previously started model at its own
+ *   discretion (for example, if a higher priority use-case which cannot coexist with detection
+ *   has been requested). The HAL must notify the framework of the preemption by sending a
+ *   recognition event with an `ABORT` status. The implementation must NOT attempt to restart the
+ *   recognition automatically when conditions change.
+ * - The HAL implementation is free to preemptively unload a previously loaded model at its own
+ *   discretion (for example, if a higher-priority model is being loaded and the two cannot
+ *   coexist). When doing so, it must first abort the detection if active (as per above) and then
+ *   notify the framework of the unload using the newly added modelUnloaded callback.
+ * - When conditions change, such that a model that couldn't previously load or start or that had
+ *   previously been preemptively stopped or unloaded, the HAL must notify the framework via the
+ *   newly added tryAgain() callback. This callback is not a guarantee that any operation would now
+ *   succeed, but merely a hint that retrying something that had previously failed, now MAY succeed.
+ *   Until this callback arrives, the framework may assume that any operation that had previously
+ *   failed or aborted would still fail if retried, so the implementation should not forget to
+ *   deliver it. There are no guarantees regarding how the framework may respond to this event and
+ *   the order in which it may choose to reload/restart its models. Typically, as result of this
+ *   event the framework will make a single attempt per model to bring this model to its desired
+ *   state (loaded, started).
+ */
+interface ISoundTriggerHw extends @2.3::ISoundTriggerHw {
+    /**
+     * This will get called at most once per every attachment to the service.
+     *
+     * All events not tied to a specific model should go through this callback.
+     */
+    registerGlobalCallback(ISoundTriggerHwGlobalCallback callback);
+
+    /**
+     * Load a sound model. Once loaded, recognition of this model can be
+     * started and stopped.
+     * The implementation returns a unique handle used by other functions
+     * (unloadSoundModel(), startRecognition*(), etc...
+     *
+     * Must have the exact same semantics as loadSoundModel from ISoundTriggerHw@2.3 except that the
+     * return values have changed and that there is no cookie provided (the implementation may pass
+     * any value to the callback, as it is ignored).
+     *
+     * @param soundModel A SoundModel structure describing the sound model
+     *     to load.
+     * @param callback The callback interface on which the soundModelCallback*()
+     *     method must be called upon completion and modelUnloaded() upon preempted unload.
+     * @return retval Operation completion status: 0 in case of success,
+     *     -EBUSY in case the operation is temporarily unavailable (but possible in general).
+     * @return modelHandle A unique handle assigned by the HAL for use by the
+     *     framework when controlling activity for this sound model.
+     */
+    loadSoundModel_2_4(SoundModel soundModel, ISoundTriggerHwCallback callback)
+            generates (int32_t retval, SoundModelHandle modelHandle);
+
+    /**
+     * Load a key phrase sound model. Once loaded, recognition of this model can
+     * be started and stopped. Only one active recognition per model at a time.
+     * The SoundTrigger service must handle concurrent recognition requests by
+     * different users/applications on the same model.
+     * The implementation returns a unique handle used by other functions
+     * (unloadSoundModel(), startRecognition*(), etc...
+     *
+     * Must have the exact same semantics as loadPhraseSoundModel from ISoundTriggerHw@2.3 except
+     * that the return values have changed and that there is no cookie provided (the implementation
+     * may pass any value to the callback, as it is ignored).
+     *
+     * @param soundModel A PhraseSoundModel structure describing the sound model
+     *     to load.
+     * @param callback The callback interface on which the soundModelCallback*()
+     *     method must be called upon completion and modelUnloaded() upon preempted unload.
+     * @return retval Operation completion status: 0 in case of success,
+     *     -EBUSY in case the operation is temporarily unavailable (but possible in general).
+     * @return modelHandle A unique handle assigned by the HAL for use by the
+     *     framework when controlling activity for this sound model.
+     */
+    loadPhraseSoundModel_2_4(PhraseSoundModel soundModel, ISoundTriggerHwCallback callback)
+            generates (int32_t retval, SoundModelHandle modelHandle);
+
+    /**
+     * Start recognition on a given model. Only one recognition active
+     * at a time per model. Once recognition succeeds or fails, the callback
+     * associated with the model handle is called.
+     *
+     * Must have the exact same semantics as startRecognition from ISoundTriggerHw@2.3 except that
+     * there are different expectations of the return value and that there is no cookie provided
+     * (the implementation may pass any value to the callback, as it is ignored).
+     *
+     * @param modelHandle the handle of the sound model to use for recognition
+     * @param config A RecognitionConfig structure containing attributes of the
+     *     recognition to perform
+     * @param callback The callback interface on which the recognitionCallback()
+     *     method must be called upon recognition.
+     * @return retval Operation completion status: 0 in case of success,
+     *     -EBUSY in case the operation is temporarily unavailable (but possible in general), or in
+     *            case model has been preemtively unloaded.
+     */
+    startRecognition_2_4(SoundModelHandle modelHandle, RecognitionConfig config)
+            generates (int32_t retval);
+};