csharp

C# Export API Example

This directory contains a modernized C# example for the current exported s2.cpp library API.

Files

• S2ExportApiExample.csproj: minimal .NET console project
• Program.cs: P/Invoke bindings for every current exported function plus sample flows

Modes

The sample supports five modes:

• smoke: validates that the exported symbols can be loaded and basic allocators work
• from-files: initializes the pipeline directly with InitializeS2PipelineFromFiles() and runs one-shot synthesis
• modular: mirrors the original SubSpecs design by creating tokenizer/model/codec separately, syncing tokenizer config from the model, binding them into a reusable pipeline, optionally precomputing prompt codes, and synthesizing
• legacy-stream: demonstrates S2SynthesizeStreaming()
• stream-ex: demonstrates S2SynthesizeStreamingEx() with the current extended options such as low latency, sentence segmentation, and .s2voice

Example Commands

One-shot modular flow close to the original idea:

dotnet run --project examples/csharp/S2ExportApiExample.csproj -- \
  --mode modular \
  --library build-vulkan/libs2.so \
  --model /path/to/s2-pro-q2_K.gguf \
  --tokenizer /path/to/tokenizer.json \
  --backend vulkan \
  --gpu-device 0 \
  --threads 8 \
  --reference-audio /path/to/reference.wav \
  --reference-text "Transcript of the reference audio." \
  --text "Hello from the modular C# example." \
  --output /tmp/s2_csharp_modular.wav

Streaming with the current extended API, using a saved .s2voice profile:

dotnet run --project examples/csharp/S2ExportApiExample.csproj -- \
  --mode stream-ex \
  --library build-vulkan/libs2.so \
  --model /path/to/s2-pro-q2_K.gguf \
  --tokenizer /path/to/tokenizer.json \
  --backend vulkan \
  --gpu-device 0 \
  --threads 8 \
  --text "First sentence. Second, longer sentence to confirm the pause." \
  --segment-sentences \
  --sentence-pause-ms 180 \
  --voice /path/to/voices/hope.s2voice \
  --output /tmp/s2_csharp_stream.wav \
  --log-level info

Comparison To The Original SubSpecs Idea

What stayed the same:

• The example still treats Pipeline, Tokenizer, Model, AudioCodec, prompt codes, and audio buffers as opaque handles owned by the native library.
• The modular flow is still first-class: load the components separately, synchronize tokenizer config from the model, bind them into a pipeline, and reuse precomputed prompt codes.
• The C# side still maps well to DllImport and pointer-style FFI usage rather than requiring a pure-C header redesign.

What changed in the updated example:

• The sample now binds every current export, not just the original one-shot subset.
• Error handling reflects the current return codes instead of the older -2/-3/-5 assumptions from the early sample.
• A runtime library-path resolver replaced the original hardcoded DLL path, so the example can load libs2.so, s2.dll, or libs2.dylib from a command-line argument.
• InitializeS2PipelineFromFiles() is covered as a convenience path in addition to the original modular path.
• The new example covers S2SynthesizeStreaming() and S2SynthesizeStreamingEx(), including low-latency streaming, sentence segmentation, and saved .s2voice loading.

What is still intentionally not "pure C":

• Handles are still modeled as opaque native objects owned by the C++ library, which matches the original intent.
• The example is designed for practical C# / DllImport interop, not for compiling the public header as strict C.