community_models: VibeASR ternary I2_S LM decoder and asr family (4/4)

CMakeLists.txt

@@ -1636,6 +1636,18 @@ audiocpp_add_model(firered_audio
         engine::models::firered_audio::make_firered_audio_loader
 )
 
+audiocpp_add_model(vibeasr
+    SOURCES
+        src/community_models/vibeasr/assets.cpp
+        src/community_models/vibeasr/vae_encoder.cpp
+        src/community_models/vibeasr/lm_decoder.cpp
+        src/community_models/vibeasr/session.cpp
+    INCLUDES
+        engine/community_models/vibeasr/session.h
+    LOADERS
+        engine::community_models::vibeasr::make_vibeasr_loader
+)
+
 set(AUDIOCPP_ENABLED_MODELS "")
 if (AUDIOCPP_MODEL_SET STREQUAL "full")
     set(AUDIOCPP_ENABLED_MODELS ${AUDIOCPP_MODEL_TARGETS})
@@ -2279,6 +2291,20 @@ if (ENGINE_BUILD_TESTS)
         COMMAND conv_lowering_matrix_test
     )
 
+    add_engine_unittest(i8_s_fused_ops_test tests/unittests/test_i8_s_fused_ops.cpp)
+
+    add_test(
+        NAME i8_s_fused_ops_test
+        COMMAND i8_s_fused_ops_test
+    )
+
+    add_engine_unittest(i2_s_mul_mat_test tests/unittests/test_i2_s_mul_mat.cpp)
+
+    add_test(
+        NAME i2_s_mul_mat_test
+        COMMAND i2_s_mul_mat_test
+    )
+
     add_engine_unittest(gguf_tensor_source_test tests/unittests/test_gguf_tensor_source.cpp)
     target_include_directories(gguf_tensor_source_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tests/unittests)
 
@@ -2418,6 +2444,56 @@ if (ENGINE_BUILD_TESTS)
             target_link_libraries(test_granite5asr_golden_transcription PRIVATE OpenMP::OpenMP_CXX)
         endif()
 
+        if (vibeasr IN_LIST AUDIOCPP_LINKED_MODELS)
+            add_executable(test_vibeasr_vae_encoder
+                tests/vibeasr/test_vibeasr_vae_encoder.cpp
+            )
+            target_compile_definitions(test_vibeasr_vae_encoder PRIVATE
+                ENGINE_REPO_ROOT="${CMAKE_CURRENT_SOURCE_DIR}"
+            )
+            target_link_libraries(test_vibeasr_vae_encoder PRIVATE engine_runtime ggml)
+            target_include_directories(test_vibeasr_vae_encoder PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+            if (ENGINE_ENABLE_OPENMP)
+                target_link_libraries(test_vibeasr_vae_encoder PRIVATE OpenMP::OpenMP_CXX)
+            endif()
+            add_test(
+                NAME test_vibeasr_vae_encoder
+                COMMAND test_vibeasr_vae_encoder
+                    --model ${CMAKE_CURRENT_SOURCE_DIR}/models/vibeasr/vibeasr-vae-encoder-i8_s.gguf
+                    --audio ${CMAKE_CURRENT_SOURCE_DIR}/assets/asr_validation/librispeech/librispeech_test_clean_6930-75918-0000.wav
+            )
+            # Needs the converted 703 MB encoder package, which a normal checkout
+            # does not have; the probe exits 125 (skip) instead of failing. Pass
+            # --reference-acoustic / --reference-semantic by hand to also check
+            # parity against a VibeASR.cpp dump.
+            set_tests_properties(test_vibeasr_vae_encoder PROPERTIES
+                SKIP_RETURN_CODE 125
+                TIMEOUT 300
+            )
+
+            add_executable(test_vibeasr_asr
+                tests/vibeasr/test_vibeasr_asr.cpp
+            )
+            target_compile_definitions(test_vibeasr_asr PRIVATE
+                ENGINE_REPO_ROOT="${CMAKE_CURRENT_SOURCE_DIR}"
+            )
+            target_link_libraries(test_vibeasr_asr PRIVATE engine_runtime ggml)
+            target_include_directories(test_vibeasr_asr PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+            if (ENGINE_ENABLE_OPENMP)
+                target_link_libraries(test_vibeasr_asr PRIVATE OpenMP::OpenMP_CXX)
+            endif()
+            add_test(
+                NAME test_vibeasr_asr
+                COMMAND test_vibeasr_asr --threads 8
+            )
+            # Same story as the encoder probe, plus the 993 MB decoder: exits 125
+            # (skip) unless both converted GGUFs sit in models/vibeasr/.
+            set_tests_properties(test_vibeasr_asr PROPERTIES
+                SKIP_RETURN_CODE 125
+                TIMEOUT 600
+            )
+        endif()
+
         if (audio8_asr IN_LIST AUDIOCPP_LINKED_MODELS)
             add_executable(test_audio8_asr_units
                 tests/audio8_asr/test_audio8_asr_units.cpp

README.md

@@ -167,6 +167,7 @@ Community model ports live under `community_models` to make the ownership bounda
 | **sense_asr** | ASR | auto, zh, en, yue, ja, ko, pt, ru, es, it, fr, de, nl, pl, tr, ar, hi, vi, th, id, ms, fa, nospeech | GGUF Q8, Stream | Jason Chen [@jasonchen31](https://github.com/jasonchen31), [@LauraGPT](https://github.com/LauraGPT) / FunASR | [SenseVoice-Small](docs/community_models/sense_asr.md) offline/streaming SAN-M + CTC transcription with event/emotion/language tags and ITN |
 | **soprano_tts** | TTS | en | GGUF Q8, Stream | [@WalkingCat](https://github.com/WalkingCat) | [Soprano-1.1-80M](https://huggingface.co/WalkingCat/Soprano-1.1-80M-GGUF) ultra-lightweight TTS with Qwen3 LM + Vocos decoder |
 | **vietneu_tts** | TTS, Clone | vi, en | GGUF | Phuoc [@phuocnguyen90](https://github.com/phuocnguyen90) | [VieNeu-TTS-v3-Turbo](docs/community_models/vietneu_tts.md) TTS and voice cloning support |
+| **vibeasr** | ASR (encoder only) | n/a | GGUF I8_S | [@XsquirrelC](https://github.com/XsquirrelC) | [VibeASR VAE encoder](docs/community_models/vibeasr.md) INT8-weight and INT8-activation port of the VibeVoice acoustic/semantic tokenizers from [VibeASR.cpp](https://github.com/microsoft/VibeASR.cpp); encoder only, the ternary I2_S decoder is not ported yet |
 | **voxcpm1** | TTS, Clone | zh, en, ja, ko | GGUF Q8, Stream | [@jasonchen31](https://github.com/jasonchen31) | [VoxCPM1](docs/community_models/voxcpm1.md) tokenizer-free 0.5B TTS with 16 kHz output, streaming, and continuation-mode voice cloning |
 
 ## Docker

docs/asr.md

@@ -329,6 +329,11 @@ chunking, server usage, and validation notes.
 
 VibeVoice ASR is an offline ASR model with greedy, sampling, and beam-search decode paths. It can return transcription text and structured segment/speaker-turn output when the model produces timestamps.
 
+A fully quantized port of the same model — INT8 activations through the encoder,
+ternary BitNet weights in the decoder — lives under community models as
+`vibeasr`: see [VibeASR](community_models/vibeasr.md). It is not a separate
+model, only a CPU-only alternative numeric pipeline for the same weights.
+
 | Field | Value |
 |---|---|
 | Family | `vibevoice_asr` |

docs/community_models/models.md

@@ -34,3 +34,4 @@ Practical expectations:
 | **sense_asr** | ASR | auto, zh, en, yue, ja, ko, pt, ru, es, it, fr, de, nl, pl, tr, ar, hi, vi, th, id, ms, fa, nospeech | Jason Chen [@jasonchen31](https://github.com/jasonchen31), [@LauraGPT](https://github.com/LauraGPT) / FunASR | [SenseVoice-Small](sense_asr.md) offline/streaming SAN-M + CTC transcription with event/emotion/language tags and ITN |
 | **vietneu_tts** | TTS, voice cloning | vi, en | Phuoc [@phuocnguyen90](https://github.com/phuocnguyen90) | [VieNeu-TTS-v3-Turbo](vietneu_tts.md) TTS and voice cloning support |
 | **moss_voicegen** | Voice design | en, zh | Joost [@jrohde](https://github.com/jrohde) | [MOSS-VoiceGenerator](moss_voicegen.md) voice design from a written instruction, on the MOSS delay architecture |
+| **vibeasr** | ASR | en | [@XsquirrelC](https://github.com/XsquirrelC) | [VibeASR](vibeasr.md) fully quantized port of [VibeASR.cpp](https://github.com/microsoft/VibeASR.cpp): the VibeVoice acoustic/semantic tokenizers on INT8 weights *and* INT8 activations through the fused `GGML_TYPE_I8_S` ops, feeding a ternary `GGML_TYPE_I2_S` BitNet Qwen2 decoder. Offline, CPU only |