Add NVIDIA Nemotron-3.5-ASR streaming multilingual backend

BENCHMARK_RESULTS.md

@@ -97,6 +97,36 @@ Comprehensive benchmark results for eddy ASR on LibriSpeech test-clean and FLEUR
 
 ---
 
+## Nemotron Streaming Multilingual 0.6B (FLEURS)
+
+**Model**: `nemotron-streaming-int8` (weight-only INT8 encoder, FP16 decoder/joint)
+**Device**: Intel NPU · **Software**: OpenVINO 2025.0 · **Decoding**: greedy, forced language
+**Preprocessor**: native C++ log-mel featurizer (replaces the dynamic-shape OV preprocessor)
+**Scoring**: FluidAudio methodology — WER for spaced languages, character-level CER for CJK
+(`ja`/`zh`), Whisper-style punctuation/symbol stripping
+
+| Language | Metric | eddy NPU (int8) | OpenVINO FP32 ref | FluidAudio CoreML ref | RTFx | Samples |
+|----------|--------|----------------:|------------------:|----------------------:|-----:|--------:|
+| English (US) | WER | 12.48 | 11.78 | 12.09 | 22.3× | 350 |
+| Spanish (LatAm) | WER | 7.03 | 6.99 | 9.01 | 22.5× | 350 |
+| French (France) | WER | 13.35 | 12.92 | 15.18 | 21.7× | 350 |
+| Chinese (Mandarin) | CER | 20.18 | 21.05 | 24.54 | 23.7× | 945 |
+| Japanese | CER | 15.46 | 15.12 | 16.86 | 23.3× | 650 |
+
+**Audio-weighted RTFx**: ~22–24× on Intel NPU (≈2× the CPU figure of ~11×).
+
+**Notes**:
+- INT8-on-NPU accuracy matches the OpenVINO FP32 reference within noise and beats the
+  FluidAudio CoreML reference on es/fr/zh/ja.
+- **NPU enablement**: the OpenVINO NPU plugin miscompiles `BitwiseNot` on a boolean
+  (integer complement → mask all-true → encoder collapses to ~0 → empty transcripts).
+  eddy rewrites `BitwiseNot → LogicalNot` in the encoder IR before compiling (no-op on
+  CPU/GPU); without it the NPU produces empty output for this model.
+- Unlike Parakeet (overlapping-chunk + 2D dedup), Nemotron is cache-aware streaming RNNT
+  with per-chunk `prompt_id` language conditioning.
+
+---
+
 ## Performance Notes
 
 ### Best Performing Languages (WER < 10%)

CMakeLists.txt

@@ -61,6 +61,10 @@ target_sources(eddy
     src/models/parakeet-v2/parakeet_decoder.cpp
     src/models/parakeet-v2/parakeet_chunking.cpp
     src/models/parakeet-v2/tokenizer.cpp
+
+    # Nemotron cache-aware streaming implementation
+    src/models/nemotron/nemotron_openvino.cpp
+    src/models/nemotron/nemotron_featurizer.cpp
 )
 
 # Link dependencies

claude.md

@@ -28,12 +28,17 @@ Match FluidAudio's Parakeet v2 (Swift/CoreML) implementation in C++/OpenVINO:
 
 ## Architecture
 
-**Batch Chunking** (current focus):
+Two first-class ASR paths are supported:
+
+**Batch (Parakeet)** — `eddy::parakeet`, stateless overlapping-chunk encoder:
 - 10s chunks with 3s overlap
 - 2D search deduplication at boundaries
 - LSTM state continuity across chunks
 
-**NOT building streaming yet** - focus is on batch processing of complete audio files.
+**Streaming (Nemotron)** — `eddy::nemotron`, cache-aware streaming FastConformer-RNNT:
+- Carries `cache_channel`/`cache_time`/`cache_len` across chunks (`att_context=[56,0]`)
+- Integer `prompt_id` per-chunk language conditioning (40+ languages)
+- Plain RNNT greedy decode (no TDT duration bins)
 
 ## Testing
 

examples/cpp/CMakeLists.txt

@@ -7,10 +7,16 @@ endif()
 add_executable(parakeet_cli parakeet_cli.cpp)
 target_link_libraries(parakeet_cli PRIVATE eddy)
 
+add_executable(nemotron_cli nemotron_cli.cpp)
+target_link_libraries(nemotron_cli PRIVATE eddy)
+
 add_executable(hf_fetch_models hf_fetch_models.cpp)
 target_link_libraries(hf_fetch_models PRIVATE eddy)
 
 add_executable(benchmark_fleurs benchmark_fleurs.cpp)
 target_link_libraries(benchmark_fleurs PRIVATE eddy)
 
-install(TARGETS parakeet_cli hf_fetch_models benchmark_fleurs DESTINATION bin)
+add_executable(benchmark_nemotron_fleurs benchmark_nemotron_fleurs.cpp)
+target_link_libraries(benchmark_nemotron_fleurs PRIVATE eddy)
+
+install(TARGETS parakeet_cli nemotron_cli hf_fetch_models benchmark_fleurs benchmark_nemotron_fleurs DESTINATION bin)

examples/cpp/benchmark_fleurs.cpp

@@ -507,7 +507,7 @@ int main(int argc, char* argv[]) {
         // Load Parakeet v3 models
         auto cache_model_dir = eddy::get_model_assets_dir("parakeet-v3");
         std::string fetch_err;
-        if (!eddy::parakeet::check_models_available(cache_model_dir, &fetch_err)) {
+        if (!eddy::model_utils::check_models_available(cache_model_dir, &fetch_err)) {
             if (!fetch_err.empty()) std::cout << "[INFO] " << fetch_err << "\n";
         }