Pre-release fixes

CMakeLists.txt

@@ -8,15 +8,17 @@ cmake_minimum_required(VERSION 3.26)
 # nemo_speech_tts_version) stay in sync without a hardcoded literal.
 file(STRINGS "${CMAKE_CURRENT_SOURCE_DIR}/VERSION" _nemo_speech_version_lines)
 foreach(_line ${_nemo_speech_version_lines})
-    if(_line MATCHES "^NEMO_SPEECH_VERSION:[ \t]*([0-9]+\\.[0-9]+\\.[0-9]+)")
-        set(NEMO_SPEECH_VERSION "${CMAKE_MATCH_1}")
+    if(_line MATCHES
+       "^NEMO_SPEECH_VERSION:[ \t]*([0-9]+\\.[0-9]+\\.[0-9]+)([-+][0-9A-Za-z.-]+)?[ \t]*$")
+        set(NEMO_SPEECH_VERSION "${CMAKE_MATCH_1}${CMAKE_MATCH_2}")
+        set(NEMO_SPEECH_PROJECT_VERSION "${CMAKE_MATCH_1}")
     endif()
 endforeach()
-if(NOT NEMO_SPEECH_VERSION)
+if(NOT NEMO_SPEECH_VERSION OR NOT NEMO_SPEECH_PROJECT_VERSION)
     message(FATAL_ERROR "could not parse NEMO_SPEECH_VERSION from ${CMAKE_CURRENT_SOURCE_DIR}/VERSION")
 endif()
 
-project(nemo_speech VERSION ${NEMO_SPEECH_VERSION} LANGUAGES C CXX)
+project(nemo_speech VERSION ${NEMO_SPEECH_PROJECT_VERSION} LANGUAGES C CXX)
 add_compile_definitions(NEMO_SPEECH_VERSION_STR="${NEMO_SPEECH_VERSION}")
 
 include(GNUInstallDirs)
@@ -35,6 +37,14 @@ install(DIRECTORY docs/
     DESTINATION "${CMAKE_INSTALL_DATADIR}/doc/nemo-speech/docs")
 install(DIRECTORY config/
     DESTINATION "${CMAKE_INSTALL_DATADIR}/nemo-speech/config")
+file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/share/nemo-speech")
+configure_file(
+    models/index.json
+    "${CMAKE_BINARY_DIR}/share/nemo-speech/model-index.json"
+    COPYONLY)
+install(FILES models/index.json
+    DESTINATION "${CMAKE_INSTALL_DATADIR}/nemo-speech"
+    RENAME model-index.json)
 
 # Windows: stop <windows.h> - pulled in transitively by the CUDA headers when
 # GGML_CUDA=ON - from defining the min()/max() macros, which otherwise clobber
@@ -93,6 +103,7 @@ option(NEMO_SPEECH_BUILD_DIAR      "Build standalone and ASR-integrated diarizat
 option(NEMO_SPEECH_BUILD_TTS       "Build text-to-speech" ON)
 option(NEMO_SPEECH_BUILD_NMT       "Build text translation (links llama.cpp)" ${NEMO_SPEECH_WITH_NMT})
 option(NEMO_SPEECH_BUILD_CLI       "Build the unified nemo-speech CLI" ON)
+option(NEMO_SPEECH_BUILD_MIC_CAPTURE "Build microphone capture in the CLI and examples" ON)
 option(NEMO_SPEECH_BUILD_HTTP      "Build the HTTP server and local playground" OFF)
 option(NEMO_SPEECH_HTTP_TLS        "Enable TLS support in the HTTP server (requires OpenSSL)" OFF)
 option(NEMO_SPEECH_BUILD_GRPC      "Build Riva-compatible gRPC adapters" ${NEMO_SPEECH_WITH_GRPC})
@@ -153,15 +164,13 @@ if(NEMO_SPEECH_BUILD_DIAR AND NOT NEMO_SPEECH_BUILD_ASR)
         "without the transcribe CLI/API surface")
 endif()
 
-# Drop-in cuBLAS shim (native GEMM, no cuBLASLt). When ON, builds
-# libcublas.so.13 from kernels/cublas_shim.cu. The shipping container image
-# substitutes it for real cuBLAS to drop ~564 MB; put it on LD_LIBRARY_PATH
-# ahead of the system cuBLAS to reproduce that GEMM path natively.
+# Drop-in cuBLAS shim (native GEMM, no cuBLASLt). Release builds can substitute
+# it for real cuBLAS to reduce their runtime closure.
 # Disabled by default for source builds, which link the CUDA toolkit's cuBLAS.
-# Container builds enable the shim explicitly to reduce the runtime image size.
+# Portable builds enable the shim explicitly.
 # It is only built when GGML_CUDA is also ON (see the target below), and is a
 # no-op for Metal, Vulkan, and CPU builds.
-option(NEMO_SPEECH_CUBLAS_SHIM     "Build the in-tree drop-in cuBLAS shim (libcublas.so.13, native GEMM, no cuBLASLt)" OFF)
+option(NEMO_SPEECH_CUBLAS_SHIM     "Build the in-tree drop-in cuBLAS shim (native GEMM, no cuBLASLt)" OFF)
 
 # Whether the linked ggml has the project ASR patches applied (ggml-patches/:
 # the fused rel-pos attention op and the F16 depthwise-conv kernel). The ASR
@@ -279,18 +288,16 @@ if(GGML_METAL)
         LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR})
 endif()
 
-# In-tree cuBLAS shim: a drop-in libcublas.so.13 backed by native GEMM kernels
-# (no cuBLASLt). ggml-cuda links real cuBLAS at build time, but the shim shares
-# its SONAME, so binaries resolve it instead when it is first on LD_LIBRARY_PATH.
-# The container images ship it in place of real cuBLAS (~564 MB saved); this
-# target reproduces that for native builds.
-# The cuBLAS shim is a Linux-only container size optimization. It relies on ELF
-# SONAME versioning (SOVERSION 13 -> libcublas.so.13) plus a GNU-ld
-# --version-script, neither of which exists with MSVC/link.exe. On Windows,
-# ggml-cuda links the real cuBLAS DLL from the CUDA toolkit, so the shim is
-# neither needed nor buildable - skip it.
-if(GGML_CUDA AND NEMO_SPEECH_CUBLAS_SHIM AND NOT WIN32)
+# In-tree cuBLAS shim backed by native GEMM kernels (no cuBLASLt). ggml-cuda
+# links real cuBLAS at build time. At runtime the loader resolves its cuBLAS
+# dependency to this target's matching major-version library name.
+if(GGML_CUDA AND NEMO_SPEECH_CUBLAS_SHIM)
     enable_language(CUDA)
+    if(CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 13.0)
+        set(NEMO_SPEECH_CUBLAS_SOVERSION 12)
+    else()
+        set(NEMO_SPEECH_CUBLAS_SOVERSION 13)
+    endif()
     add_library(nemo_speech_cublas_shim SHARED kernels/cublas_shim.cu)
     if(CMAKE_CUDA_ARCHITECTURES)
         set(NEMO_SPEECH_CUBLAS_SHIM_ARCHITECTURES "${CMAKE_CUDA_ARCHITECTURES}")
@@ -300,16 +307,25 @@ if(GGML_CUDA AND NEMO_SPEECH_CUBLAS_SHIM AND NOT WIN32)
         set(NEMO_SPEECH_CUBLAS_SHIM_ARCHITECTURES "80-virtual")
     endif()
     set_target_properties(nemo_speech_cublas_shim PROPERTIES
-        OUTPUT_NAME cublas
-        SOVERSION 13
         CUDA_ARCHITECTURES "${NEMO_SPEECH_CUBLAS_SHIM_ARCHITECTURES}"
+        CUDA_RUNTIME_LIBRARY Static
         CUDA_SEPARABLE_COMPILATION ON)
-    target_link_options(nemo_speech_cublas_shim PRIVATE
-        "LINKER:--version-script=${CMAKE_CURRENT_SOURCE_DIR}/kernels/ver_cublas.map")
-elseif(GGML_CUDA AND NEMO_SPEECH_CUBLAS_SHIM AND WIN32)
-    message(STATUS
-        "NEMO_SPEECH_CUBLAS_SHIM: skipped on Windows; linking the CUDA "
-        "toolkit's real cuBLAS instead.")
+    if(WIN32)
+        # The toolkit import library records this exact DLL name. App-local
+        # deployment therefore substitutes the shim without changing ggml.
+        set_target_properties(nemo_speech_cublas_shim PROPERTIES
+            OUTPUT_NAME "cublas64_${NEMO_SPEECH_CUBLAS_SOVERSION}")
+    else()
+        set_target_properties(nemo_speech_cublas_shim PROPERTIES
+            OUTPUT_NAME cublas
+            SOVERSION "${NEMO_SPEECH_CUBLAS_SOVERSION}")
+        configure_file(
+            kernels/ver_cublas.map
+            "${CMAKE_CURRENT_BINARY_DIR}/ver_cublas.map"
+            @ONLY)
+        target_link_options(nemo_speech_cublas_shim PRIVATE
+            "LINKER:--version-script=${CMAKE_CURRENT_BINARY_DIR}/ver_cublas.map")
+    endif()
 endif()
 
 if(NEMO_SPEECH_WITH_FLASHLIGHT)
@@ -448,6 +464,10 @@ if(DEFINED VCPKG_INSTALLED_DIR AND DEFINED VCPKG_TARGET_TRIPLET)
 endif()
 install(FILES ggml/LICENSE
     DESTINATION "${NEMO_SPEECH_THIRD_PARTY_LICENSE_DIR}/ggml")
+if(NEMO_SPEECH_BUILD_ASR AND NEMO_SPEECH_BUILD_CLI AND NEMO_SPEECH_BUILD_MIC_CAPTURE)
+    install(FILES third_party/miniaudio/LICENSE
+        DESTINATION "${NEMO_SPEECH_THIRD_PARTY_LICENSE_DIR}/miniaudio")
+endif()
 if(TARGET nemo_speech_cublas_shim)
     install(TARGETS nemo_speech_cublas_shim
         RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
@@ -500,7 +520,7 @@ configure_package_config_file(
 write_basic_package_version_file(
     ${CMAKE_CURRENT_BINARY_DIR}/NeMoSpeechConfigVersion.cmake
     VERSION ${PROJECT_VERSION}
-    COMPATIBILITY SameMajorVersion)
+    COMPATIBILITY SameMinorVersion)
 install(FILES
     ${CMAKE_CURRENT_BINARY_DIR}/NeMoSpeechConfig.cmake
     ${CMAKE_CURRENT_BINARY_DIR}/NeMoSpeechConfigVersion.cmake

CMakePresets.json

@@ -206,49 +206,53 @@
     {
       "name": "cpu-server",
       "inherits": "cpu-asr",
-      "displayName": "CPU ASR and TTS HTTP server and playground",
+      "displayName": "CPU speech CLI, HTTP server, and playground",
       "cacheVariables": {
         "NEMO_SPEECH_BUILD_TTS": "ON",
+        "NEMO_SPEECH_BUILD_NMT": "ON",
         "NEMO_SPEECH_BUILD_HTTP": "ON",
         "NEMO_SPEECH_BUILD_GRPC": "OFF",
         "NEMO_SPEECH_WITH_GRPC": "OFF",
-        "NEMO_SPEECH_WITH_NMT": "OFF"
+        "NEMO_SPEECH_WITH_NMT": "ON"
       }
     },
     {
       "name": "cuda-server",
       "inherits": "cuda-asr",
-      "displayName": "CUDA ASR and TTS HTTP server and playground",
+      "displayName": "CUDA speech CLI, HTTP server, and playground",
       "cacheVariables": {
         "NEMO_SPEECH_BUILD_TTS": "ON",
+        "NEMO_SPEECH_BUILD_NMT": "ON",
         "NEMO_SPEECH_BUILD_HTTP": "ON",
         "NEMO_SPEECH_BUILD_GRPC": "OFF",
         "NEMO_SPEECH_WITH_GRPC": "OFF",
-        "NEMO_SPEECH_WITH_NMT": "OFF"
+        "NEMO_SPEECH_WITH_NMT": "ON"
       }
     },
     {
       "name": "metal-server",
       "inherits": "metal-asr",
-      "displayName": "Metal ASR and TTS HTTP server and playground",
+      "displayName": "Metal speech CLI, HTTP server, and playground",
       "cacheVariables": {
         "NEMO_SPEECH_BUILD_TTS": "ON",
+        "NEMO_SPEECH_BUILD_NMT": "ON",
         "NEMO_SPEECH_BUILD_HTTP": "ON",
         "NEMO_SPEECH_BUILD_GRPC": "OFF",
         "NEMO_SPEECH_WITH_GRPC": "OFF",
-        "NEMO_SPEECH_WITH_NMT": "OFF"
+        "NEMO_SPEECH_WITH_NMT": "ON"
       }
     },
     {
       "name": "vulkan-server",
       "inherits": "vulkan-asr",
-      "displayName": "Vulkan ASR and TTS HTTP server and playground",
+      "displayName": "Vulkan speech CLI, HTTP server, and playground",
       "cacheVariables": {
         "NEMO_SPEECH_BUILD_TTS": "ON",
+        "NEMO_SPEECH_BUILD_NMT": "ON",
         "NEMO_SPEECH_BUILD_HTTP": "ON",
         "NEMO_SPEECH_BUILD_GRPC": "OFF",
         "NEMO_SPEECH_WITH_GRPC": "OFF",
-        "NEMO_SPEECH_WITH_NMT": "OFF"
+        "NEMO_SPEECH_WITH_NMT": "ON"
       }
     },
     {

CONTRIBUTING.md

@@ -2,6 +2,29 @@
 
 We welcome external contributions to NeMo-Speech.cpp.
 
+## Development checks
+
+Follow the [source-build guide](docs/build.md) for prerequisites and submodules.
+For a model-independent CPU ASR test build:
+
+```bash
+git submodule update --init ggml llama.cpp
+scripts/configure.sh cpu-asr -DNEMO_SPEECH_BUILD_TESTS=ON
+cmake --build --preset cpu-asr
+ctest --test-dir build/cpu-asr --output-on-failure
+```
+
+Install [pre-commit](https://pre-commit.com/) and run the same formatting,
+license-header, and static file checks used by CI:
+
+```bash
+pre-commit run --all-files
+```
+
+Use the closest matching CUDA, Metal, Vulkan, server, or component preset when
+the change affects code outside the CPU ASR path. Include the commands and
+results relevant to the change in the pull request.
+
 ## Contribution license and provenance
 
 Unless a file states otherwise, contributions are submitted under the

README.md

@@ -1,9 +1,22 @@
 # NeMo-Speech.cpp
 
-A lightweight native C++ runtime for NVIDIA Nemotron Speech models built on ggml. Runs speech models in realtime and in batch mode across platforms/backends.
+A lightweight native C++ runtime for running the NVIDIA Nemotron Speech model family locally, with broad hardware support. It supports multilingual speech recognition, speaker diarization, translation, and speech synthesis in real-time and batch modes.
+
+NeMo-Speech.cpp is NVIDIA's official local speech inference solution, with day-0 support for our latest speech models. It builds on models from [NVIDIA NeMo Speech](https://github.com/NVIDIA-NeMo/Speech), with native inference powered by [ggml](https://github.com/ggml-org/ggml).
+
+## Models and applications
+
+| Application | Supported models |
+|---|---|
+| Speech recognition | [Nemotron 3.5 ASR Streaming 0.6B](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b), [Nemotron Speech Streaming 0.6B](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b), [Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3), and [Parakeet CTC 1.1B](https://huggingface.co/nvidia/parakeet-ctc-1.1b) |
+| Speaker diarization | [Streaming Sortformer 4-speaker v2](https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2), standalone or combined with ASR |
+| Text and speech translation | [Riva Translate 4B Instruct v2](https://huggingface.co/nvidia/Riva-Translate-4B-Instruct-v2), with composed ASR-to-NMT-to-TTS speech translation |
+| Speech synthesis | [MagpieTTS Multilingual 357M](https://huggingface.co/nvidia/magpie_tts_multilingual_357m) with [NeMo NanoCodec](https://huggingface.co/nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps) |
+| Speech processing | [Silero VAD](https://github.com/snakers4/silero-vad), punctuation and capitalization, endpointing, text normalization, and subtitles |
 
 ## Contents
 
+- [Models and applications](#models-and-applications)
 - [Installation](#installation)
 - [Quick start](#quick-start)
 - [Command line](#command-line)
@@ -16,38 +29,57 @@ A lightweight native C++ runtime for NVIDIA Nemotron Speech models built on ggml
 
 ## Installation
 
-From a source checkout, install the CLI, HTTP API, and browser playground for
-the detected platform and backend:
+Install the `nemo-speech` CLI for the detected platform and backend:
+
+On Linux or macOS, run:
 
 ```bash
-scripts/install.sh --source
+curl -fsSL https://github.com/NVIDIA/NeMo-Speech.cpp/raw/main/scripts/install.sh | sh
 export PATH="$HOME/.local/bin:$PATH"  # current shell; future shells are updated
 ```
 
-The source build requires Git, CMake 3.26 or newer, Ninja, a C++17 compiler,
-and the toolkit for the selected GPU backend. See
-[Installation](docs/install.md) for platform-specific prerequisites, options,
-and the native release-archive flow.
+On Windows, run from PowerShell:
+
+```powershell
+irm https://github.com/NVIDIA/NeMo-Speech.cpp/raw/main/scripts/install.ps1 | iex
+```
+
+Open a new PowerShell window after installation so the updated user `PATH`
+takes effect.
+
+The installer prefers a verified native release and falls back to a source
+build when an artifact is unavailable. A source build requires Git, CMake 3.26
+or newer, Ninja, a C++17 compiler, SentencePiece development files, and the
+toolchain required by the selected backend, if any. See
+[Installation](docs/install.md) for platform-specific prerequisites and
+options.
 
 ## Quick start
 
-Download the ready-to-run Q8 GGUF from the model's Hugging Face repository,
-then transcribe the bundled sample:
+Transcribe a local WAV file. On first use, the CLI downloads the pinned default
+Nemotron 3.5 GGUF from Hugging Face and verifies its size and SHA-256:
 
 ```bash
-hf download nvidia/nemotron-speech-streaming-en-0.6b \
-  nemotron-speech-streaming-en-0.6b.q8_0.gguf \
-  --local-dir models
+nemo-speech transcribe /path/to/audio.wav
+```
 
-nemo-speech transcribe test_files/asr/wav/test/jfk.wav \
-  --model models/nemotron-speech-streaming-en-0.6b.q8_0.gguf
+Source checkouts can use `test_files/asr/wav/test/jfk.wav` as a smoke-test
+input.
+
+The same command can transcribe the default microphone on builds that include
+live capture:
+
+```bash
+nemo-speech transcribe --live
 ```
 
-Install the `hf` command with `pip install -U huggingface_hub` if needed. The
-CLI selects an available backend and handles common mono or stereo PCM WAV
-sample rates automatically. Substitute your own WAV file after verifying the
-bundled sample. See [ASR models](docs/asr/models.md) for the other published
-GGUFs and [model conversion](docs/model-conversion.md) for custom checkpoints.
+Run `nemo-speech model list` to see defaults, short names, and which command
+uses each model. For example, `nemo-speech pull nemotron-en` downloads the
+English-only model ahead of time, and `--model nemotron-en` selects it. Local
+GGUF paths continue to work without downloading anything. The CLI selects an
+available backend and handles common mono or stereo PCM WAV sample rates
+automatically. See the [CLI model guide](docs/cli.md#models-and-cache) and
+[model conversion](docs/model-conversion.md) for custom checkpoints.
 
 ## Command line
 
@@ -62,13 +94,13 @@ Start the same runtime as a local HTTP service and open the playground:
 
 ```bash
 nemo-speech serve \
-  --asr-model models/nemotron-speech-streaming-en-0.6b.q8_0.gguf \
+  --asr-model nemotron-3.5 \
   --open
 ```
 
-The server binds to <http://127.0.0.1:8080> by default and also provides a
-documented OpenAI-compatible audio API subset and realtime WebSocket
-transcription. A separately built `riva_server` binary provides the
+The server binds to <http://127.0.0.1:8080> by default. Its transcription and
+speech routes expose documented OpenAI-compatible subsets, alongside realtime
+WebSocket transcription. A separately built `riva_server` binary provides the
 Riva-compatible gRPC interface. See the [server guide](docs/server.md) when you
 are ready to integrate either interface.
 
@@ -88,11 +120,12 @@ gRPC usage.
 
 ## Build from source
 
-Requires CMake 3.26 or newer, Ninja, C and C++17 compilers, and a supported
-CUDA toolkit. For a CUDA ASR and TTS server with the playground:
+Requires CMake 3.26 or newer, Ninja, C and C++17 compilers, SentencePiece
+development files, and the toolchain required by the selected backend, if any.
+For a CUDA ASR and TTS server with the playground:
 
 ```bash
-git submodule update --init ggml third_party/cpp-httplib
+git submodule update --init ggml llama.cpp third_party/cpp-httplib
 scripts/configure.sh cuda-server
 cmake --build --preset cuda-server
 ```
@@ -119,9 +152,13 @@ Windows, and container instructions are in
 
 ## License
 
-NVIDIA-authored code is released under the [Apache License 2.0](LICENSE), with
-the project copyright notice in [NOTICE](NOTICE). Third-party components retain
-their respective terms; see [Third-Party Notices](THIRD_PARTY_NOTICES.md).
+NVIDIA-authored code is released under the
+[Apache License 2.0](https://github.com/NVIDIA/NeMo-Speech.cpp/blob/main/LICENSE),
+with the project copyright notice in
+[NOTICE](https://github.com/NVIDIA/NeMo-Speech.cpp/blob/main/NOTICE). Third-party
+components retain their respective terms; see
+[Third-Party Notices](https://github.com/NVIDIA/NeMo-Speech.cpp/blob/main/THIRD_PARTY_NOTICES.md).
+Release archives also include these files under `share/licenses/nemo-speech/`.
 
 ## Contributing