feat(models): add IBM Granite Speech 5.0 470M TurboCTC ASR engine

.gitignore

@@ -19,6 +19,7 @@ __pycache__/
 /tests/omnivoice/outputs/
 /unittests/
 /models/
+/granite5asr/
 /reference/
 /resources/
 /patches/

CMakeLists.txt

@@ -1271,6 +1271,22 @@ audiocpp_add_model(citrinet_asr
         engine::models::citrinet_asr::make_citrinet_asr_loader
 )
 
+audiocpp_add_model(granite5asr
+    SOURCES
+        src/community_models/granite5asr/assets.cpp
+        src/community_models/granite5asr/frontend.cpp
+        src/community_models/granite5asr/encoder.cpp
+        src/community_models/granite5asr/session.cpp
+    INCLUDES
+        engine/community_models/granite5asr/session.h
+    LOADERS
+        engine::community_models::granite5asr::make_granite5asr_loader
+    ALIASES
+        granite_speech5_asr
+        granite_speech
+        granite_speech5_ctc
+)
+
 audiocpp_add_model(vevo2
     SOURCES
         src/models/vevo2/ar.cpp
@@ -1801,6 +1817,31 @@ add_executable(miocodec_wavlm_parity EXCLUDE_FROM_ALL
 
 target_link_libraries(miocodec_wavlm_parity PRIVATE engine_runtime ggml)
 
+add_executable(test_granite5asr_golden_transcription
+    tests/granite5asr/test_granite5asr_golden_transcription.cpp
+)
+target_compile_definitions(test_granite5asr_golden_transcription PRIVATE
+    ENGINE_REPO_ROOT="${CMAKE_CURRENT_SOURCE_DIR}"
+)
+target_link_libraries(test_granite5asr_golden_transcription PRIVATE engine_runtime ggml)
+target_include_directories(test_granite5asr_golden_transcription PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+if (ENGINE_ENABLE_OPENMP)
+    target_link_libraries(test_granite5asr_golden_transcription PRIVATE OpenMP::OpenMP_CXX)
+endif()
+
+add_executable(granite5asr_warm_bench
+    tests/granite5asr/granite5asr_warm_bench.cpp
+)
+target_compile_definitions(granite5asr_warm_bench PRIVATE
+    ENGINE_REPO_ROOT="${CMAKE_CURRENT_SOURCE_DIR}"
+)
+target_link_libraries(granite5asr_warm_bench PRIVATE engine_runtime ggml)
+target_include_directories(granite5asr_warm_bench PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+if (ENGINE_ENABLE_OPENMP)
+    target_link_libraries(granite5asr_warm_bench PRIVATE OpenMP::OpenMP_CXX)
+endif()
+
+
 # F5/Habibi tests: parity harnesses + e2e sample generator. Only when the
 # f5_tts model is linked (they call model-internal symbols) AND test hooks
 # are explicitly enabled — F5_MEL_TEST must never leak into production
@@ -1930,6 +1971,7 @@ if (ENGINE_BUILD_WARMBENCH)
     add_engine_warmbench(confucius4_tts_warm_bench tests/confucius4_tts/confucius4_tts_warm_bench.cpp)
     add_engine_warmbench(controlfoley_warm_bench tests/controlfoley/controlfoley_warm_bench.cpp)
     add_engine_warmbench(fun_asr_nano_warm_bench tests/fun_asr_nano/fun_asr_nano_warm_bench.cpp)
+    add_engine_warmbench(granite5asr_warm_bench tests/granite5asr/granite5asr_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)

docs/community_models/granite5asr.md

@@ -0,0 +1,38 @@
+# IBM Granite Speech 5.0 470M TurboCTC in audio.cpp
+
+IBM Granite Speech 5.0 TurboCTC is a compact 470-million-parameter English Automatic Speech Recognition (ASR) model delivering state-of-the-art transcription accuracy with ultra-low latency.
+
+## Architecture
+
+- **Audio Frontend**: 16 kHz, 80-bin HTK mel spectrogram with 8.0 dB dynamic flooring, first-order deltas, and 2x frame stacking (320-dim features).
+- **Acoustic Conformer**: 16 layers ($d_{model}=1024, d_{ff}=4096, d_{conv}=2048, heads=8, head\_dim=128$) with block self-attention ($context\_size=128$), Shaw relative positional embeddings, depthwise convolution with folded batch-norm, and mid-layer CTC self-conditioning at layer 8.
+- **Decoder**: Non-autoregressive Connectionist Temporal Classification (CTC) with greedy decoding and 16,384 BPE vocabulary via fast HuggingFace tokenizer.
+
+## CLI Usage
+
+### Offline Transcription
+
+```bash
+# Transcribe audio using native safetensors directory
+audiocpp_cli asr --model granite5asr --input sample.wav
+
+# Or explicitly pass the family
+audiocpp_cli asr --family granite5asr --model path/to/checkpoint --input sample.wav
+```
+
+### Quantized GGUF Loading
+
+```bash
+# Transcribe using a converted Q8_0 GGUF package
+audiocpp_cli asr --model models/granite-speech-5.0-470m-turboctc-gguf --input sample.wav
+```
+
+### Audio Chunking & VAD
+
+```bash
+# Long audio with automatic VAD segmentation
+audiocpp_cli asr --model granite5asr --input long_audio.wav --option audio_chunk_mode=auto
+
+# Fixed 30-second chunking
+audiocpp_cli asr --model granite5asr --input long_audio.wav --option audio_chunk_mode=fixed --option audio_chunk_duration_sec=30
+```

docs/community_models/models.md

@@ -18,6 +18,7 @@ Practical expectations:
 |---|---|---|---|---|
 | **f5_tts** | TTS, voice cloning | en, ar (Habibi) | Community | [F5-TTS](f5_tts.md) flow-matching DiT — M0 scaffolding, aliases `habibi`/`habibi_tts` |
 | **glm_tts** | TTS, voice cloning | zh, en | Mirek [@mirek190](https://github.com/mirek190) | [GLM-TTS](glm_tts.md) zero-shot synthesis and voice cloning support |
+| **granite5asr** | ASR | en | Community | [IBM Granite Speech 5.0 470M TurboCTC](granite5asr.md) ultra-fast Conformer-CTC ASR with Shaw relative positional embeddings and ByteLevel BPE |
 | **inflect_v2** | TTS | en | Community | [Inflect Micro v2 and Nano v2](inflect_v2.md) native FP32 offline synthesis |
 | **kroko_asr** | ASR | de, en, es, fr, it, he, nl, pt, sv, tr | Mirek [@mirek190](https://github.com/mirek190) | [Kroko Community ASR](kroko_asr.md) native offline/streaming Zipformer2/RNN-T transcription with word timestamps |
 | **mms_forced_aligner** | Align | nl (nld), en (eng); pre-romanized Latin | Community | [MMS-300M-1130 Forced Aligner](mms_forced_aligner.md) word-timestamp alignment from a wav2vec2 CTC checkpoint (safetensors or local GGUF) |

include/engine/community_models/granite5asr/assets.h

@@ -0,0 +1,66 @@
+#pragma once
+
+#include "engine/framework/assets/tensor_source.h"
+#include "engine/framework/model_spec/package.h"
+#include "engine/framework/tokenizers/hf_tokenizer_json.h"
+
+#include <cstdint>
+#include <filesystem>
+#include <memory>
+#include <string>
+#include <vector>
+
+namespace engine::community_models::granite5asr {
+
+struct Granite5FrontendConfig {
+    int64_t sample_rate = 16000;
+    int64_t n_fft = 512;
+    int64_t win_length = 400;
+    int64_t hop_length = 160;
+    int64_t n_mels = 80;
+    int64_t stack_factor = 2;
+    bool deltas = true;
+    int64_t delta_win_length = 3;
+    float logmel_floor_db = 8.0f;
+};
+
+struct Granite5EncoderConfig {
+    int64_t hidden_size = 1024;
+    int64_t intermediate_size = 4096;
+    int64_t num_layers = 16;
+    int64_t num_attention_heads = 8;
+    int64_t num_key_value_heads = 8;
+    int64_t head_dim = 128;
+    int64_t context_size = 128;
+    int64_t conv_kernel_size = 7;
+    int64_t conv_expansion_factor = 2;
+    int64_t max_position_embeddings = 512;
+    int64_t num_mel_bins = 80;
+    int64_t input_features = 320;
+    int64_t vocab_size = 16384;
+    std::vector<int64_t> subsample_layers = {0, 1};
+};
+
+struct Granite5ASRConfig {
+    std::string model_type = "granite_speech5_ctc";
+    int64_t vocab_size = 16384;
+    int64_t pad_token_id = 0;
+    int64_t blank_token_id = 0;
+    Granite5FrontendConfig frontend;
+    Granite5EncoderConfig encoder;
+};
+
+struct Granite5ASRAssets {
+    assets::ResourceBundle resources;
+    std::shared_ptr<const assets::TensorSource> source;
+    Granite5ASRConfig config;
+    std::shared_ptr<tokenizers::HuggingFaceTokenizerJson> tokenizer;
+    std::vector<uint8_t> special_token_ids;
+};
+
+Granite5ASRConfig parse_granite5asr_config(const std::string & json_text);
+
+std::shared_ptr<const Granite5ASRAssets> load_granite5asr_assets(
+    const std::filesystem::path & model_path);
+
+}  // namespace engine::community_models::granite5asr

include/engine/community_models/granite5asr/encoder.h

@@ -0,0 +1,72 @@
+#pragma once
+
+#include "engine/framework/core/backend.h"
+#include "engine/framework/core/backend_weight_store.h"
+#include "engine/framework/core/execution_context.h"
+#include "engine/framework/modules/linear_module.h"
+#include "engine/framework/modules/norm_modules.h"
+#include "engine/community_models/granite5asr/assets.h"
+#include "engine/community_models/granite5asr/frontend.h"
+
+#include <cstdint>
+#include <memory>
+#include <vector>
+
+namespace engine::community_models::granite5asr {
+
+struct Granite5LayerWeights {
+    modules::NormWeights ffn1_norm;
+    modules::LinearWeights ffn1_fc1;
+    modules::LinearWeights ffn1_fc2;
+
+    modules::NormWeights norm_self_att;
+    core::TensorValue q_proj;
+    core::TensorValue k_proj;
+    core::TensorValue v_proj;
+    modules::LinearWeights o_proj;
+    core::TensorValue rel_pos_emb;
+
+    modules::NormWeights norm_conv;
+    modules::LinearWeights conv_pw1;
+    core::TensorValue conv_dw_weight;
+    core::TensorValue conv_dw_bias;
+    modules::LinearWeights conv_pw2;
+
+    modules::NormWeights ffn2_norm;
+    modules::LinearWeights ffn2_fc1;
+    modules::LinearWeights ffn2_fc2;
+
+    modules::NormWeights norm_out;
+
+    bool is_subsample = false;
+};
+
+struct Granite5EncoderWeights {
+    modules::LinearWeights input_linear;
+    std::vector<Granite5LayerWeights> layers;
+    modules::LinearWeights out;
+    modules::LinearWeights out_mid;
+};
+
+class Granite5EncoderRuntime {
+public:
+    Granite5EncoderRuntime(
+        std::shared_ptr<const Granite5ASRAssets> assets,
+        engine::core::ExecutionContext & execution_context,
+        assets::TensorStorageType storage_type,
+        size_t graph_arena_bytes = 1024ull * 1024ull * 1024ull);
+
+    std::vector<int32_t> transcribe_features(
+        const Granite5FrontendFeatures & features);
+
+    const Granite5ASRAssets & assets() const noexcept { return *assets_; }
+
+private:
+    std::shared_ptr<const Granite5ASRAssets> assets_;
+    engine::core::ExecutionContext * execution_context_ = nullptr;
+    engine::core::BackendWeightStore weight_store_;
+    Granite5EncoderWeights weights_;
+    size_t graph_arena_bytes_;
+};
+
+}  // namespace engine::community_models::granite5asr

include/engine/community_models/granite5asr/frontend.h

@@ -0,0 +1,34 @@
+#pragma once
+
+#include "engine/framework/audio/dsp.h"
+#include "engine/framework/runtime/session.h"
+#include "engine/community_models/granite5asr/assets.h"
+
+#include <memory>
+#include <vector>
+
+namespace engine::community_models::granite5asr {
+
+struct Granite5FrontendFeatures {
+    std::vector<float> values;
+    int64_t frames = 0;
+    int64_t feature_dim = 320;
+};
+
+class Granite5Frontend {
+public:
+    explicit Granite5Frontend(std::shared_ptr<const Granite5ASRAssets> assets);
+
+    Granite5FrontendFeatures extract(const runtime::AudioBuffer & audio) const;
+    Granite5FrontendFeatures extract_waveform(const std::vector<float> & waveform) const;
+    std::vector<float> prepare_waveform(const runtime::AudioBuffer & audio) const;
+
+    const Granite5FrontendConfig & config() const noexcept { return assets_->config.frontend; }
+
+private:
+    std::shared_ptr<const Granite5ASRAssets> assets_;
+    audio::AudioTensor mel_filterbank_;
+    std::vector<float> window_;
+};
+
+}  // namespace engine::community_models::granite5asr