Expand description
Local speech-to-text inference built around Candle.
The crate exposes a single API for model asset resolution, runtime selection, audio decoding, and transcription across the supported ASR backends. Model weights can be supplied from directories, explicit files, typed in-memory assets, or the optional Hugging Face downloader.
Re-exports§
pub use audio::AudioInput;pub use audio::AudioSamples;pub use config::DType;pub use config::ModelAsset;pub use config::ModelAssetBundle;pub use config::ModelAssetDir;pub use config::ModelFiles;pub use config::MoonshineAssets;pub use config::Qwen3AsrAssets;pub use config::Qwen3TokenizerAsset;pub use config::RuntimeChoice;pub use config::SafetensorsWeights;pub use config::SttConfig;pub use config::WhisperAssets;pub use config::preferred_runtime_choice;pub use config::preferred_runtime_choices;pub use device::DeviceSelection;pub use error::SttError;pub use models::BackendKind;pub use models::InferenceMode;pub use models::ModelAssetRequirement;pub use models::ModelType;pub use traits::ModelInfo;pub use traits::SttModel;pub use traits::TimestampGranularity;pub use traits::TranscriptionRequest;pub use traits::TranscriptionResult;pub use traits::TranscriptionSegment;pub use traits::TranscriptionTask;
Modules§
- audio
- config
- device
- Device selection utilities.
- download
- error
- models
- opus
- Opus packet decoding backed by a vendored, statically built libopus.
- traits
Functions§
- load_
model - Load a model backend based on the provided configuration.