Add Nemotron-ASR streaming inference to C++ SDK

sdk/cpp/CMakeLists.txt

@@ -54,6 +54,8 @@ add_library(CppSdk STATIC
   src/catalog.cpp
   src/openai_chat_client.cpp
   src/openai_audio_client.cpp
+  src/openai_live_audio_types.cpp
+  src/openai_live_audio_client.cpp
   src/foundry_local_manager.cpp
 )
 
@@ -91,6 +93,7 @@ if (BUILD_TESTING)
     test/model_variant_test.cpp
     test/catalog_test.cpp
     test/client_test.cpp
+    test/live_audio_test.cpp
   )
 
   target_include_directories(CppSdkTests

sdk/cpp/include/foundry_local.h

@@ -16,3 +16,5 @@
 #include "openai/openai_tool_types.h"
 #include "openai/openai_chat_client.h"
 #include "openai/openai_audio_client.h"
+#include "openai/openai_live_audio_types.h"
+#include "openai/openai_live_audio_client.h"

sdk/cpp/include/openai/openai_audio_client.h

@@ -7,6 +7,7 @@
 #include <string_view>
 #include <functional>
 #include <filesystem>
+#include <memory>
 
 #include <gsl/pointers>
 
@@ -22,6 +23,8 @@ namespace foundry_local {
         std::string text;
     };
 
+    class LiveAudioTranscriptionSession;
+
     class OpenAIAudioClient final {
     public:
         explicit OpenAIAudioClient(const IModel& model);
@@ -34,6 +37,9 @@ namespace foundry_local {
         using StreamCallback = std::function<void(const AudioCreateTranscriptionResponse& chunk)>;
         void TranscribeAudioStreaming(const std::filesystem::path& audioFilePath, const StreamCallback& onChunk) const;
 
+        /// Create a new live audio transcription session for streaming PCM audio.
+        std::unique_ptr<LiveAudioTranscriptionSession> CreateLiveTranscriptionSession() const;
+
     private:
         OpenAIAudioClient(gsl::not_null<foundry_local::Internal::IFoundryLocalCore*> core, std::string_view modelId,
                           gsl::not_null<ILogger*> logger);

sdk/cpp/include/openai/openai_live_audio_client.h

@@ -0,0 +1,101 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+
+#include <string>
+#include <memory>
+#include <thread>
+#include <mutex>
+#include <atomic>
+#include <chrono>
+#include <cstdint>
+#include <vector>
+
+#include <gsl/pointers>
+
+#include "openai_live_audio_types.h"
+
+namespace foundry_local::Internal {
+    struct IFoundryLocalCore;
+    template <typename T> class ThreadSafeQueue;
+} // namespace foundry_local::Internal
+
+namespace foundry_local {
+    class ILogger;
+
+    class LiveAudioTranscriptionSession final {
+    public:
+        LiveAudioTranscriptionSession(gsl::not_null<Internal::IFoundryLocalCore*> core,
+                                       std::string modelId,
+                                       gsl::not_null<ILogger*> logger);
+        ~LiveAudioTranscriptionSession() noexcept;
+
+        // Non-copyable, non-movable
+        LiveAudioTranscriptionSession(const LiveAudioTranscriptionSession&) = delete;
+        LiveAudioTranscriptionSession& operator=(const LiveAudioTranscriptionSession&) = delete;
+        LiveAudioTranscriptionSession(LiveAudioTranscriptionSession&&) = delete;
+        LiveAudioTranscriptionSession& operator=(LiveAudioTranscriptionSession&&) = delete;
+
+        /// Mutable settings reference; only effective before Start().
+        LiveAudioTranscriptionOptions& Settings() { return settings_; }
+        /// Read-only settings reference.
+        const LiveAudioTranscriptionOptions& Settings() const { return settings_; }
+        /// Settings that were active when Start() was called.
+        const LiveAudioTranscriptionOptions& ActiveSettings() const { return activeSettings_; }
+
+        /// Begin the streaming session. Must be called before Append/TryAppend.
+        void Start();
+
+        /// Enqueue PCM audio data. Blocks if the push queue is full.
+        void Append(const uint8_t* pcmData, size_t length);
+
+        /// Try to get the next transcription result within the given timeout.
+        TranscriptionStatus TryGetNext(LiveAudioTranscriptionResponse& result,
+                                        std::chrono::milliseconds timeout = std::chrono::seconds(5));
+
+        /// Signal the end of audio input and stop the session.
+        void Stop();
+
+        /// Returns the error message if the session is in an error state.
+        std::string GetErrorMessage() const;
+
+        /// Returns true if the session has been started.
+        bool IsStarted() const;
+
+        /// Returns true if the session has been stopped.
+        bool IsStopped() const;
+
+    private:
+        enum class SessionState {
+            Created,
+            Starting,
+            Started,
+            Stopped
+        };
+
+        void PushWorkerLoop();
+        void StopInternal(std::unique_lock<std::mutex>& lock);
+
+        gsl::not_null<Internal::IFoundryLocalCore*> core_;
+        std::string modelId_;
+        gsl::not_null<ILogger*> logger_;
+
+        LiveAudioTranscriptionOptions settings_;
+        LiveAudioTranscriptionOptions activeSettings_;
+
+        mutable std::mutex mutex_;
+        SessionState state_ = SessionState::Created;
+        std::string sessionHandle_;
+
+        using AudioChunk = std::vector<uint8_t>;
+        std::unique_ptr<Internal::ThreadSafeQueue<AudioChunk>> pushQueue_;
+        std::unique_ptr<Internal::ThreadSafeQueue<LiveAudioTranscriptionResponse>> resultQueue_;
+
+        std::thread pushThread_;
+        std::string errorMessage_;
+        LiveAudioTranscriptionResponse finalResult_;
+        bool hasFinalResult_ = false;
+    };
+
+} // namespace foundry_local

sdk/cpp/include/openai/openai_live_audio_types.h

@@ -0,0 +1,50 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include <optional>
+
+namespace foundry_local {
+
+    struct ContentPart {
+        std::string text;
+        std::string transcript;
+    };
+
+    struct LiveAudioTranscriptionResponse {
+        std::string text;
+        bool is_final = false;
+        std::optional<double> start_time;
+        std::optional<double> end_time;
+        std::vector<ContentPart> content;
+
+        static LiveAudioTranscriptionResponse FromJson(const std::string& json);
+    };
+
+    struct LiveAudioTranscriptionOptions {
+        int sample_rate = 16000;
+        int channels = 1;
+        int bits_per_sample = 16;
+        std::optional<std::string> language;
+        int push_queue_capacity = 100;
+    };
+
+    struct CoreErrorResponse {
+        std::string code;
+        std::string message;
+        bool is_transient = false;
+
+        static std::optional<CoreErrorResponse> TryParse(const std::string& error_string);
+    };
+
+    enum class TranscriptionStatus {
+        Result,
+        Timeout,
+        Closed,
+        Error
+    };
+
+} // namespace foundry_local

sdk/cpp/src/core.h

@@ -1,7 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 //
-// Core DLL interop – loads Microsoft.AI.Foundry.Local.Core.dll at runtime.
+// Core DLL interop � loads Microsoft.AI.Foundry.Local.Core.dll at runtime.
 // Internal header, not part of the public API.
 
 #pragma once
@@ -46,6 +46,7 @@ namespace foundry_local {
             module_.reset();
             execCmd_ = nullptr;
             execCbCmd_ = nullptr;
+            execBinaryCmd_ = nullptr;
             freeResCmd_ = nullptr;
         }
 
@@ -91,10 +92,55 @@ namespace foundry_local {
             return result;
         }
 
+        CoreResponse callWithBinary(std::string_view command, ILogger& logger,
+                                     const std::string* dataArgument,
+                                     const uint8_t* binaryData, size_t binaryDataLength) const override {
+            if (!module_ || !execBinaryCmd_ || !freeResCmd_) {
+                throw Exception("Core is not loaded. Cannot call command: " + std::string(command), logger);
+            }
+
+            StreamingRequestBuffer request{};
+            request.Command = command.empty() ? nullptr : command.data();
+            request.CommandLength = static_cast<int32_t>(command.size());
+
+            if (dataArgument && !dataArgument->empty()) {
+                request.Data = dataArgument->data();
+                request.DataLength = static_cast<int32_t>(dataArgument->size());
+            }
+
+            if (binaryData && binaryDataLength > 0) {
+                if (binaryDataLength > static_cast<size_t>(INT32_MAX)) {
+                    throw Exception("Binary data length exceeds maximum supported size (INT32_MAX).", logger);
+                }
+                request.BinaryData = binaryData;
+                request.BinaryDataLength = static_cast<int32_t>(binaryDataLength);
+            }
+
+            ResponseBuffer response{};
+            auto safeDeleter = [fn = freeResCmd_](ResponseBuffer* buf) {
+                if (fn)
+                    fn(buf);
+            };
+            std::unique_ptr<ResponseBuffer, decltype(safeDeleter)> responseGuard(&response, safeDeleter);
+
+            execBinaryCmd_(&request, &response);
+
+            CoreResponse result;
+            if (response.Error && response.ErrorLength > 0) {
+                result.error.assign(static_cast<const char*>(response.Error), response.ErrorLength);
+                return result;
+            }
+            if (response.Data && response.DataLength > 0) {
+                result.data.assign(static_cast<const char*>(response.Data), response.DataLength);
+            }
+            return result;
+        }
+
     private:
         wil::unique_hmodule module_;
         execute_command_fn execCmd_{};
         execute_command_with_callback_fn execCbCmd_{};
+        execute_command_with_binary_fn execBinaryCmd_{};
         free_response_fn freeResCmd_{};
 
         void LoadFromPath(const std::filesystem::path& path) {
@@ -105,6 +151,8 @@ namespace foundry_local {
             execCmd_ = reinterpret_cast<execute_command_fn>(RequireProc(m.get(), "execute_command"));
             execCbCmd_ = reinterpret_cast<execute_command_with_callback_fn>(
                 RequireProc(m.get(), "execute_command_with_callback"));
+            execBinaryCmd_ = reinterpret_cast<execute_command_with_binary_fn>(
+                RequireProc(m.get(), "execute_command_with_binary"));
             freeResCmd_ = reinterpret_cast<free_response_fn>(RequireProc(m.get(), "free_response"));
 
             module_ = std::move(m);

sdk/cpp/src/flcore_native.h

@@ -26,14 +26,25 @@ extern "C"
     // Callback signature: void(*)(void* data, int length, void* userData)
     using UserCallbackFn = void(__cdecl*)(void*, int32_t, void*);
 
+    struct StreamingRequestBuffer {
+        const void* Command;
+        int32_t CommandLength;
+        const void* Data;
+        int32_t DataLength;
+        const void* BinaryData;
+        int32_t BinaryDataLength;
+    };
+
     // Exported function pointer types
     using execute_command_fn = void(__cdecl*)(RequestBuffer*, ResponseBuffer*);
     using execute_command_with_callback_fn = void(__cdecl*)(RequestBuffer*, ResponseBuffer*, void* /*callback*/,
                                                             void* /*userData*/);
+    using execute_command_with_binary_fn = void(__cdecl*)(StreamingRequestBuffer*, ResponseBuffer*);
     using free_response_fn = void(__cdecl*)(ResponseBuffer*);
 
     static_assert(std::is_standard_layout<RequestBuffer>::value, "RequestBuffer must be standard layout");
     static_assert(std::is_standard_layout<ResponseBuffer>::value, "ResponseBuffer must be standard layout");
+    static_assert(std::is_standard_layout<StreamingRequestBuffer>::value, "StreamingRequestBuffer must be standard layout");
 
 #pragma pack(pop)
 }

sdk/cpp/src/foundry_local_internal_core.h

@@ -31,6 +31,11 @@ namespace foundry_local {
             virtual CoreResponse call(std::string_view command, ILogger& logger,
                                       const std::string* dataArgument = nullptr, NativeCallbackFn callback = nullptr,
                                       void* data = nullptr) const = 0;
+
+            virtual CoreResponse callWithBinary(std::string_view command, ILogger& logger,
+                                                const std::string* dataArgument,
+                                                const uint8_t* binaryData, size_t binaryDataLength) const = 0;
+
             virtual void unload() = 0;
         };
 

sdk/cpp/src/openai_audio_client.cpp

@@ -16,6 +16,8 @@
 #include "core_helpers.h"
 #include "logger.h"
 
+#include "openai/openai_live_audio_client.h"
+
 namespace foundry_local {
 
     OpenAIAudioClient::OpenAIAudioClient(gsl::not_null<Internal::IFoundryLocalCore*> core, std::string_view modelId,
@@ -67,4 +69,8 @@ namespace foundry_local {
         }
     }
 
+    std::unique_ptr<LiveAudioTranscriptionSession> OpenAIAudioClient::CreateLiveTranscriptionSession() const {
+        return std::make_unique<LiveAudioTranscriptionSession>(core_, modelId_, logger_);
+    }
+
 } // namespace foundry_local