Add Fun-ASR-Nano offline ASR

CMakeLists.txt

@@ -279,6 +279,7 @@ add_library(engine_core OBJECT
     src/framework/modules/vocoders/hift_vocoder.cpp
     src/framework/modules/speech_encoders/hubert_encoder.cpp
     src/framework/modules/speech_encoders/wavlm_encoder.cpp
+    src/framework/modules/speech_encoders/sanm.cpp
     src/framework/modules/optimizations/fast_conv_modules.cpp
     src/framework/modules/optimizations/fast_projection_modules.cpp
     src/framework/modules/recurrent_modules.cpp
@@ -311,6 +312,7 @@ add_library(engine_core OBJECT
     external/llama_tokenizer/unicode-data.cpp
     src/framework/tokenizers/sentencepiece.cpp
     src/framework/audio/fft.cpp
+    src/framework/audio/kaldi_fbank.cpp
     src/framework/audio/activity.cpp
     src/framework/audio/chunking.cpp
     src/framework/audio/conversion.cpp
@@ -689,6 +691,23 @@ audiocpp_add_model(dramabox
         engine::models::dramabox::make_dramabox_loader
 )
 
+audiocpp_add_model(fun_asr_nano
+    SOURCES
+        src/models/fun_asr_nano/assets.cpp
+        src/models/fun_asr_nano/frontend.cpp
+        src/models/fun_asr_nano/encoder.cpp
+        src/models/fun_asr_nano/adaptor.cpp
+        src/models/fun_asr_nano/tokenizer_text.cpp
+        src/models/fun_asr_nano/prompt.cpp
+        src/models/fun_asr_nano/decoder.cpp
+        src/models/fun_asr_nano/session.cpp
+        src/models/fun_asr_nano/loader.cpp
+    INCLUDES
+        engine/models/fun_asr_nano/loader.h
+    LOADERS
+        engine::models::fun_asr_nano::make_fun_asr_nano_loader
+)
+
 audiocpp_add_model(heartmula
     SOURCES
         src/models/heartmula/assets.cpp
@@ -1330,6 +1349,7 @@ if (ENGINE_BUILD_WARMBENCH)
     add_engine_warmbench(citrinet_asr_warm_bench tests/citrinet_asr/citrinet_asr_warm_bench.cpp)
     add_engine_warmbench(confucius4_tts_warm_bench tests/confucius4_tts/confucius4_tts_warm_bench.cpp)
     add_engine_warmbench(dramabox_warm_bench tools/dramabox/dramabox_warm_bench.cpp)
+    add_engine_warmbench(fun_asr_nano_warm_bench tests/fun_asr_nano/fun_asr_nano_warm_bench.cpp)
     add_engine_warmbench(higgs_audio_stt_warm_bench tests/higgs_audio_stt/higgs_audio_stt_warm_bench.cpp)
     add_engine_warmbench(higgs_audio_tts_warm_bench tests/higgs_audio_tts/higgs_audio_tts_warm_bench.cpp)
     add_engine_warmbench(hviske_asr_warm_bench tests/hviske_asr/hviske_asr_warm_bench.cpp)
@@ -1666,6 +1686,214 @@ if (ENGINE_BUILD_TESTS)
         COMMAND model_spec_system_test
     )
 
+    add_engine_unittest(fun_asr_nano_assets_test tests/unittests/test_fun_asr_nano_assets.cpp)
+    target_sources(fun_asr_nano_assets_test PRIVATE src/models/fun_asr_nano/assets.cpp)
+    target_include_directories(fun_asr_nano_assets_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tests/unittests)
+
+    add_test(
+        NAME fun_asr_nano_assets_test
+        COMMAND fun_asr_nano_assets_test
+    )
+
+    add_engine_unittest(fun_asr_nano_frontend_probe tests/fun_asr_nano/fun_asr_nano_frontend_probe.cpp)
+
+    add_test(
+        NAME fun_asr_nano_frontend_probe
+        COMMAND fun_asr_nano_frontend_probe
+            ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/frontend_reference.json
+            ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/frontend_reference.bin
+            ${CMAKE_CURRENT_SOURCE_DIR}/assets/resources/sample_16k.wav
+    )
+
+    add_engine_unittest(fun_asr_nano_sanm_probe tests/fun_asr_nano/sanm_probe.cpp)
+
+    add_test(
+        NAME fun_asr_nano_sanm_probe
+        COMMAND fun_asr_nano_sanm_probe
+            --backend cpu
+            --reference ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/sanm_reference.json
+            --data ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/sanm_reference.bin
+    )
+
+    add_engine_unittest(fun_asr_nano_encoder_probe tests/fun_asr_nano/fun_asr_nano_encoder_probe.cpp)
+    target_sources(
+        fun_asr_nano_encoder_probe
+        PRIVATE
+            src/models/fun_asr_nano/assets.cpp
+            src/models/fun_asr_nano/frontend.cpp
+            src/models/fun_asr_nano/encoder.cpp
+    )
+    set(
+        AUDIOCPP_FUN_ASR_NANO_TEST_MODEL
+        ""
+        CACHE PATH
+        "Optional local Fun-ASR-Nano model directory for full runtime parity tests"
+    )
+    set(
+        AUDIOCPP_FUN_ASR_NANO_TEST_GGUF
+        ""
+        CACHE FILEPATH
+        "Optional standalone Fun-ASR-Nano GGUF for full CLI transcript tests"
+    )
+    if (AUDIOCPP_FUN_ASR_NANO_TEST_MODEL)
+        add_test(
+            NAME fun_asr_nano_encoder_probe_cpu
+            COMMAND fun_asr_nano_encoder_probe
+                --backend cpu
+                --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                --reference ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/encoder_reference.json
+                --data ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/encoder_reference.bin
+        )
+        if (GGML_CUDA)
+            add_test(
+                NAME fun_asr_nano_encoder_probe_cuda
+                COMMAND fun_asr_nano_encoder_probe
+                    --backend cuda
+                    --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                    --reference ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/encoder_reference.json
+                    --data ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/encoder_reference.bin
+            )
+        endif()
+
+        add_engine_unittest(fun_asr_nano_session_probe tests/fun_asr_nano/fun_asr_nano_session_probe.cpp)
+        add_test(
+            NAME fun_asr_nano_session_probe_cpu
+            COMMAND fun_asr_nano_session_probe
+                --backend cpu
+                --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                --audio ${CMAKE_CURRENT_SOURCE_DIR}/assets/resources/sample_16k.wav
+        )
+        if (GGML_CUDA)
+            add_test(
+                NAME fun_asr_nano_session_probe_cuda
+                COMMAND fun_asr_nano_session_probe
+                    --backend cuda
+                    --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                    --audio ${CMAKE_CURRENT_SOURCE_DIR}/assets/resources/sample_16k.wav
+            )
+        endif()
+    endif()
+    if (AUDIOCPP_FUN_ASR_NANO_TEST_GGUF)
+        add_test(
+            NAME fun_asr_nano_gguf_cli_cpu
+            COMMAND audiocpp_cli
+                --task asr
+                --family fun_asr_nano
+                --model ${AUDIOCPP_FUN_ASR_NANO_TEST_GGUF}
+                --backend cpu
+                --audio ${CMAKE_CURRENT_SOURCE_DIR}/assets/resources/sample_16k.wav
+        )
+        set_tests_properties(
+            fun_asr_nano_gguf_cli_cpu
+            PROPERTIES PASS_REGULAR_EXPRESSION
+                "text_output=Some call me nature, others call me mother nature"
+        )
+        if (GGML_CUDA)
+            add_test(
+                NAME fun_asr_nano_gguf_cli_cuda
+                COMMAND audiocpp_cli
+                    --task asr
+                    --family fun_asr_nano
+                    --model ${AUDIOCPP_FUN_ASR_NANO_TEST_GGUF}
+                    --backend cuda
+                    --audio ${CMAKE_CURRENT_SOURCE_DIR}/assets/resources/sample_16k.wav
+            )
+            set_tests_properties(
+                fun_asr_nano_gguf_cli_cuda
+                PROPERTIES PASS_REGULAR_EXPRESSION
+                    "text_output=Some call me nature, others call me mother nature"
+            )
+            add_test(
+                NAME fun_asr_nano_gguf_cli_cuda_shared_q8
+                COMMAND audiocpp_cli
+                    --task asr
+                    --family fun_asr_nano
+                    --model ${AUDIOCPP_FUN_ASR_NANO_TEST_GGUF}
+                    --backend cuda
+                    --session-option fun_asr_nano.weight_type=q8_0
+                    --audio ${CMAKE_CURRENT_SOURCE_DIR}/assets/resources/sample_16k.wav
+            )
+            set_tests_properties(
+                fun_asr_nano_gguf_cli_cuda_shared_q8
+                PROPERTIES PASS_REGULAR_EXPRESSION
+                    "text_output=Some call me nature, others call me mother nature"
+            )
+            add_test(
+                NAME fun_asr_nano_gguf_cli_best_shared_q8
+                COMMAND audiocpp_cli
+                    --task asr
+                    --family fun_asr_nano
+                    --model ${AUDIOCPP_FUN_ASR_NANO_TEST_GGUF}
+                    --backend best
+                    --session-option fun_asr_nano.weight_type=q8_0
+                    --audio ${CMAKE_CURRENT_SOURCE_DIR}/assets/resources/sample_16k.wav
+            )
+            set_tests_properties(
+                fun_asr_nano_gguf_cli_best_shared_q8
+                PROPERTIES PASS_REGULAR_EXPRESSION
+                    "text_output=Some call me nature, others call me mother nature"
+            )
+        endif()
+    endif()
+
+    add_engine_unittest(fun_asr_nano_adaptor_probe tests/fun_asr_nano/fun_asr_nano_adaptor_probe.cpp)
+    target_sources(
+        fun_asr_nano_adaptor_probe
+        PRIVATE
+            src/models/fun_asr_nano/assets.cpp
+            src/models/fun_asr_nano/adaptor.cpp
+    )
+    if (AUDIOCPP_FUN_ASR_NANO_TEST_MODEL)
+        add_test(
+            NAME fun_asr_nano_adaptor_probe_cpu
+            COMMAND fun_asr_nano_adaptor_probe
+                --backend cpu
+                --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                --reference ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/adaptor_reference.json
+                --data ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/adaptor_reference.bin
+        )
+        if (GGML_CUDA)
+            add_test(
+                NAME fun_asr_nano_adaptor_probe_cuda
+                COMMAND fun_asr_nano_adaptor_probe
+                    --backend cuda
+                    --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                    --reference ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/adaptor_reference.json
+                    --data ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/adaptor_reference.bin
+            )
+        endif()
+    endif()
+
+    add_engine_unittest(fun_asr_nano_decoder_probe tests/fun_asr_nano/fun_asr_nano_decoder_probe.cpp)
+    target_sources(
+        fun_asr_nano_decoder_probe
+        PRIVATE
+            src/models/fun_asr_nano/assets.cpp
+            src/models/fun_asr_nano/tokenizer_text.cpp
+            src/models/fun_asr_nano/prompt.cpp
+            src/models/fun_asr_nano/decoder.cpp
+    )
+    if (AUDIOCPP_FUN_ASR_NANO_TEST_MODEL)
+        add_test(
+            NAME fun_asr_nano_decoder_probe_cpu
+            COMMAND fun_asr_nano_decoder_probe
+                --backend cpu
+                --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                --reference ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/decoder_reference.json
+                --data ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/decoder_reference.bin
+        )
+        if (GGML_CUDA)
+            add_test(
+                NAME fun_asr_nano_decoder_probe_cuda
+                COMMAND fun_asr_nano_decoder_probe
+                    --backend cuda
+                    --model ${AUDIOCPP_FUN_ASR_NANO_TEST_MODEL}
+                    --reference ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/decoder_reference.json
+                    --data ${CMAKE_CURRENT_SOURCE_DIR}/tests/fun_asr_nano/decoder_reference.bin
+            )
+        endif()
+    endif()
+
     add_engine_unittest(tdt_decoder_duration_loop_test tests/unittests/test_tdt_decoder_duration_loop.cpp)
     target_include_directories(tdt_decoder_duration_loop_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tests/unittests)
 

README.md

@@ -61,6 +61,7 @@ Runtime tags: safetensors is the default model loading path. `GGUF 16/Q8/Q4` mea
 | **chatterbox** | TTS, Clone, VC| ar, da, de, el, en, es, fi, fr, hi, it, ko, ms, nl, no, pl, pt, sv, sw, tr | Chatterbox with 0.5B backbone | GGUF 16/Q8 |
 | **citrinet_asr** | ASR | en | Citrinet-256 | GGUF Q8 |
 | **fish_audio** | TTS, Clone, Ctrl | auto, en, zh | Fish Audio S2 Pro | GGUF 16/Q8 |
+| **fun_asr_nano** | ASR | auto, zh, en, ja | Fun-ASR-Nano-2512 | GGUF 16/Q8 |
 | **heartmula** | Music | zh, en, ja, ko, es | HeartMuLa-oss-3B with HeartCodec-oss | GGUF 16/Q8 |
 | **higgs_audio_stt** | ASR | en | Higgs Audio v3 STT | GGUF 16/Q8, Stream |
 | **higgs_audio_tts** | TTS, Clone, Ctrl | auto | Higgs Audio v3 TTS 4B | GGUF 16/Q8 |

docs/asr.md

@@ -2,6 +2,7 @@
 
 | Model | Family | Mode(s) | Quick Start |
 |---|---|---|---|
+| Fun-ASR-Nano | `fun_asr_nano` | offline | [Fun-ASR-Nano](#fun-asr-nano) |
 | Qwen3 ASR | `qwen3_asr` | offline | [Qwen3 ASR](#qwen3-asr) |
 | Citrinet ASR | `citrinet_asr` | offline | [Citrinet ASR](#citrinet-asr) |
 | Kroko Community ASR | `kroko_asr` | offline, streaming | [Kroko Community ASR](#kroko-community-asr) |
@@ -21,6 +22,24 @@ audiocpp_cli --task asr --family <family> --model <model-dir> --backend cuda --a
 
 When `--mode streaming` is used, the selected model provides its default streaming policy.
 
+## Fun-ASR-Nano
+
+Fun-ASR-Nano provides offline multilingual transcription for Chinese, English,
+and Japanese with automatic language selection. The recommended package is the
+standalone Q8_0 GGUF published by FunAudioLLM.
+
+```bash
+python3 tools/model_manager_v2.py install fun_asr_nano
+audiocpp_cli --task asr --family fun_asr_nano \
+  --model models/Fun-ASR-Nano-2512-GGUF/fun-asr-nano-2512-q8_0.gguf \
+  --backend cuda --audio speech_16k.wav --text-out transcript.txt
+```
+
+The runtime supports fixed offline chunking and inverse text normalization.
+Streaming and timestamp output are not exposed. See the
+[Fun-ASR-Nano model guide](models/fun_asr_nano.md) for package, option, GGUF,
+and server details.
+
 ## Qwen3 ASR
 
 Qwen3 ASR transcribes speech and can be paired with Qwen3 Forced Aligner when timestamps are needed. See [Qwen3 models](models/qwen3.md) for the full ASR and alignment manual.

docs/gguf.md

@@ -61,6 +61,7 @@ Status labels:
 | `chatterbox` | Done | Pass | --- | Pass (ASR match, drift) | Pass (ASR match, drift) |
 | `citrinet_asr` | Done | Pass | --- | --- | Pass |
 | `fish_audio` | Done | Pass | --- | Pass | Pass |
+| `fun_asr_nano` | Done | Pass | --- | Pass | Pass |
 | `glm_tts` | Done | Pass (TTS + clone) | --- | --- | Pass (ASR match, drift) |
 | `heartmula` | Done | Pass | --- | Pass (drift) | Pass (drift) |
 | `higgs_audio_stt` | Done | Pass | --- | Pass | Pass |

docs/model_manager.md

@@ -14,6 +14,7 @@ If a model has a ready-to-use GGUF package, prefer that route first.
 Ready-to-use GGUF packages are published here:
 
 - Core released models: [audio-cpp/audio.cpp-gguf](https://huggingface.co/audio-cpp/audio.cpp-gguf)
+- Fun-ASR-Nano: [FunAudioLLM/Fun-ASR-Nano-2512-GGUF](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512-GGUF)
 - Community OuteTTS package: [mirek190/audio.cpp](https://huggingface.co/mirek190/audio.cpp/tree/main/Text%20to%20audio%20(TTS))
 
 For support status and tested precision coverage, see the [GGUF guide](gguf.md).
@@ -112,6 +113,8 @@ Packages whose loaders are not registered in the current release tree are listed
 | `chatterbox` | Chatterbox | **Yes** |
 | `citrinet_asr` | Citrinet ASR converted layout | No |
 | `fish_audio_s2_pro` | Fish Audio S2 Pro GGUF Q8_0 | **Yes** |
+| `fun_asr_nano_2512_f16` | Fun-ASR-Nano-2512 F16 GGUF | **Yes** |
+| `fun_asr_nano_2512_q8_0` | Fun-ASR-Nano-2512 Q8_0 GGUF | **Yes** |
 | `glm_tts` | GLM-TTS with converted Flow/HiFT/tokenizer and CAMPPlus assets | No |
 | `heartmula` | HeartMuLa | No |
 | `higgs_audio_stt` | Higgs Audio STT | No |