speechcore

Reusable Rust speech-to-text runtime components for desktop and service applications.

speechcore contains reusable local STT building blocks: microphone capture, Silero VAD segmentation, backend selection, model provisioning, transcript streaming, and runtime state. It deliberately does not contain UI, clipboard integration, system tray code, global shortcuts, or transcript enhancement behavior.

What You Get

• Push-to-talk and real-time transcription orchestration.
• Audio capture via PortAudio.
• Silero VAD for speech segmentation.
• Pluggable transcription backends:
  • CTranslate2
  • Moonshine ONNX
  • NVIDIA Parakeet ONNX
  • NVIDIA Nemotron 3.5 ASR ONNX (streaming: cache-aware FastConformer + RNNT, with low-latency partials)
  • whisper.cpp via whisper-rs when explicitly enabled
• Model cache and download helpers.
• Transcript and backend status streams for app-owned UI or service logic.

Platform Support

Developed and CI-tested on Linux x86_64 (Ubuntu in CI, NixOS in dev). The pure-Rust API (default features, no backends) is portable, but each backend pulls native dependencies that determine real platform support:

• backend-whisper-cpp builds plain CPU whisper.cpp by default (portable). Acceleration is opt-in: whisper-cpp-vulkan needs the Vulkan SDK (glslc), whisper-cpp-openblas needs OpenBLAS.
• The ONNX backends and Silero VAD use ort, which downloads a per-platform ONNX Runtime by default; combining them with backend-ctranslate2 requires a system ONNX Runtime (see ONNX Runtime + CTranslate2).

Contributions to validate or fix other platforms are welcome.

Installing

Published on crates.io. The default crate has no heavy native runtime features enabled — opt into the runtime and the backend(s) you want:

speechcore = { version = "0.1", default-features = false, features = ["runtime", "backend-whisper-cpp"] }

To track a specific Git revision or tag instead of the crates.io release:

speechcore = { git = "https://github.com/0xPD33/speechcore", tag = "v0.1.1", default-features = false, features = ["runtime", "backend-whisper-cpp"] }

For local development from this workspace:

speechcore = {
  path = "../speechcore",
  default-features = false,
  features = ["runtime", "backend-parakeet"]
}

Feature Flags

Available features:

backend-whisper-cpp is intentionally opt-in at the library level. Apps that also link another ggml consumer, such as a llama.cpp enhancement layer, should make an explicit choice about enabling both native stacks.

Minimal Manual Session

use speechcore::prelude::*;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let mut engine = SpeechEngine::new(SpeechConfig::default()).await?;
    let _session_id = engine.start_session().await?;
    tokio::time::sleep(std::time::Duration::from_secs(3)).await;
    let transcript = engine.stop_and_transcribe().await?;
    println!("{}", transcript.text);

    engine.shutdown().await?;
    Ok(())
}

Logging

The library uses tracing for diagnostics. Applications decide how logs are collected and displayed:

tracing_subscriber::fmt::init();

Native Dependencies

Some features need native libraries:

• audio-capture: PortAudio and system audio headers.
• backend-ctranslate2: CTranslate2 dependencies required by ct2rs.
• ONNX backends and Silero VAD: ONNX Runtime as required by ort.
• backend-whisper-cpp: whisper.cpp build dependencies and selected acceleration features.

On NixOS or other declarative environments, prefer project-local shells or flakes that pin these dependencies.

ONNX Runtime + CTranslate2 (important)

ort (used by the ONNX backends and Silero VAD) downloads a static ONNX Runtime by default. That static build bundles protobuf, which collides with CTranslate2's protobuf at link time (duplicate google::protobuf::* symbols) whenever both are enabled — and runtime enables Silero VAD (ort), so this hits any runtime + backend-ctranslate2 build. Link a system ONNX Runtime so it is loaded dynamically instead:

# Install ONNX Runtime (matching ort's expected version, e.g. 1.22.x) to /usr/local, then:
export ORT_STRATEGY=system
export ORT_LIB_LOCATION=/usr/local/lib
export LIBRARY_PATH=/usr/local/lib       # link-time resolution of -lonnxruntime
export LD_LIBRARY_PATH=/usr/local/lib    # run-time loading
# OpenBLAS (whisper.cpp `openblas` feature) needs its headers located:
export BLAS_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu
export OPENBLAS_PATH=/usr

The Nix flake sets this up automatically; .github/workflows/ci.yml is a known-good Ubuntu reference for the native dependencies and env vars.

Model Cache

Model downloads are stored in ~/.cache/speechcore/models by default, or under $XDG_CACHE_HOME/speechcore/models when XDG_CACHE_HOME is set. Applications can override this with SPEECHCORE_MODEL_DIR.

Repository Checks

Recommended checks before publishing or opening a PR:

cargo fmt --all -- --check
cargo check
cargo test --no-default-features
cargo check --no-default-features --features runtime
cargo clippy --no-default-features --features runtime --all-targets -- -D warnings
cargo test --no-default-features --features runtime,backend-ctranslate2
cargo check --no-default-features --features runtime,all-backends
cargo check --no-default-features --features runtime,backend-whisper-cpp --examples
cargo doc --no-deps
cargo package
cargo publish --dry-run

When validating an uncommitted local extraction, cargo package --allow-dirty and cargo publish --dry-run --allow-dirty can preview the package contents. Use plain cargo package and cargo publish --dry-run before publishing or tagging a release.

License

Licensed under either of:

• Apache License, Version 2.0
• MIT license

at your option.