feat(model): add Confucius4-R2T2 real-time streaming ASR (r2t2_asr)

CMakeLists.txt

@@ -1469,6 +1469,22 @@ audiocpp_add_model(qwen3_asr
         qwen3_forced_aligner
 )
 
+audiocpp_add_model(confucius4_r2t2
+    SOURCES
+        src/community_models/confucius4_r2t2/assets.cpp
+        src/community_models/confucius4_r2t2/text_postprocess.cpp
+        src/community_models/confucius4_r2t2/tokenizer_text.cpp
+        src/community_models/confucius4_r2t2/frontend_whisper.cpp
+        src/community_models/confucius4_r2t2/audio_encoder.cpp
+        src/community_models/confucius4_r2t2/thinker.cpp
+        src/community_models/confucius4_r2t2/prompt_asr.cpp
+        src/community_models/confucius4_r2t2/session.cpp
+    INCLUDES
+        engine/community_models/confucius4_r2t2/session.h
+    LOADERS
+        engine::community_models::confucius4_r2t2::make_confucius4_r2t2_loader
+)
+
 audiocpp_add_model(qwen3_forced_aligner
     SOURCES
         src/models/qwen3_forced_aligner/processor.cpp
@@ -2971,6 +2987,23 @@ if (ENGINE_BUILD_TESTS OR ENGINE_BUILD_EXTENDED_TESTS OR ENGINE_BUILD_MODEL_TEST
             target_link_libraries(test_granite5asr_golden_transcription PRIVATE OpenMP::OpenMP_CXX)
         endif()
 
+        if (confucius4_r2t2 IN_LIST AUDIOCPP_LINKED_MODELS)
+            add_executable(test_confucius4_r2t2_graph_reuse tests/confucius4_r2t2/test_graph_reuse.cpp)
+            target_compile_definitions(test_confucius4_r2t2_graph_reuse PRIVATE ENGINE_REPO_ROOT="${CMAKE_CURRENT_SOURCE_DIR}")
+            target_link_libraries(test_confucius4_r2t2_graph_reuse PRIVATE engine_runtime ggml)
+            add_executable(test_confucius4_r2t2_transcription
+                tests/confucius4_r2t2/test_confucius4_r2t2_transcription.cpp
+            )
+            target_compile_definitions(test_confucius4_r2t2_transcription PRIVATE
+                ENGINE_REPO_ROOT="${CMAKE_CURRENT_SOURCE_DIR}"
+            )
+            target_link_libraries(test_confucius4_r2t2_transcription PRIVATE engine_runtime ggml)
+            target_include_directories(test_confucius4_r2t2_transcription PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+            if (ENGINE_ENABLE_OPENMP)
+                target_link_libraries(test_confucius4_r2t2_transcription PRIVATE OpenMP::OpenMP_CXX)
+            endif()
+        endif()
+
         if (vibeasr IN_LIST AUDIOCPP_LINKED_MODELS)
             add_executable(test_vibeasr_vae_encoder
                 tests/vibeasr/test_vibeasr_vae_encoder.cpp

app/server/example.json

@@ -36,6 +36,24 @@
       "path": "../../models/Qwen3-ASR-0.6B",
       "task": "asr",
       "mode": "offline"
+    },
+    {
+      "id": "r2t2-asr",
+      "family": "confucius4_r2t2",
+      "path": "../../models/Confucius4-R2T2-GGUF/r2t2-q8_0.gguf",
+      "task": "asr",
+      "mode": "offline"
+    },
+    {
+      "id": "r2t2-asr-stream",
+      "family": "confucius4_r2t2",
+      "path": "../../models/Confucius4-R2T2-GGUF/r2t2-q8_0.gguf",
+      "task": "asr",
+      "mode": "streaming",
+      "session_options": {
+        "confucius4_r2t2.chunk_size_ms": "320",
+        "confucius4_r2t2.max_tokens": "32"
+      }
     }
   ]
 }

docs/asr.md

@@ -7,6 +7,7 @@
 | Fun-ASR-Nano | `fun_asr_nano` | offline | [Fun-ASR-Nano](#fun-asr-nano) |
 | Granite Speech 5.0 TurboCTC | `granite5asr` | offline | [Granite Speech 5.0 TurboCTC](community_models/granite5asr.md) |
 | Qwen3 ASR | `qwen3_asr` | offline, streaming | [Qwen3 ASR](#qwen3-asr) |
+| Confucius4-R2T2 | `confucius4_r2t2` | offline, streaming | [Confucius4-R2T2](community_models/r2t2.md) |
 | Citrinet ASR | `citrinet_asr` | offline | [Citrinet ASR](#citrinet-asr) |
 | Kroko Community ASR | `kroko_asr` | offline, streaming | [Kroko Community ASR](#kroko-community-asr) |
 | Higgs Audio STT | `higgs_audio_stt` | offline, streaming | [Higgs Audio STT](models/higgs_audio_stt.md) |
@@ -61,6 +62,28 @@ audiocpp_cli --task asr --family qwen3_asr --model models/Qwen3-ASR-1.7B-hf --ba
 audiocpp_cli --task asr --mode streaming --family qwen3_asr --model models/Qwen3-ASR-1.7B-hf --backend cuda --audio speech_16k.wav --request-option audio_chunk_seconds=5 --text-out transcript.txt
 ```
 
+## Confucius4-R2T2
+
+Confucius4-R2T2 is a low-latency append-only streaming ASR model: a Qwen3-ASR
+1.7B fine-tune with Longest Stable Prefix (LSP) decoding. Committed text is
+never revised, and chunk sizes from 80 ms to 2 s are supported. It runs the
+same audio tower as Qwen3 ASR, so only the streaming state machine differs.
+
+```bash
+audiocpp_cli --task asr --family confucius4_r2t2 --model models/Confucius4-R2T2 \
+  --backend metal --audio speech_16k.wav --text-out transcript.txt
+```
+
+```bash
+audiocpp_cli --task asr --mode streaming --family confucius4_r2t2 \
+  --model models/Confucius4-R2T2 --backend metal --audio speech_16k.wav \
+  --session-option confucius4_r2t2.chunk_size_ms=320 --text-out transcript.txt
+```
+
+Streaming emits append-only partial text; the final transcript is returned when
+the stream ends. See the [Confucius4-R2T2 model guide](community_models/r2t2.md) for the
+session options, the LSP state machine, and the MPS golden verification recipe.
+
 ## Citrinet ASR
 
 Citrinet is an offline CTC ASR model. It produces transcription text from speech audio.