Crate yamabiko_whisper 

Streaming speech recognition on top of whisper-rs, using the LocalAgreement-2 policy from Macháček et al. 2023.

whisper-rs is fully encapsulated: callers do not need to add it to their own Cargo.toml or import any of its types. Build a processor with OnlineAsrModel::create_processor or OnlineAsrProcessor::from_model_path, feed it 16 kHz mono f32 PCM via OnlineAsrProcessor::insert_audio_chunk, or let AsrPipeline handle downmixing, resampling, and chunking for a microphone/file/network source.

Accelerated whisper.cpp backends are exposed as Cargo features (cuda, vulkan, metal, coreml, hipblas, intel-sycl, openblas, openmp). Use BackendConfig when loading a model to select a GPU device or force CPU execution.

Structs

AsrPipeline

High-level audio-to-ASR pipeline.

AudioInputConfig

Source audio format accepted by AsrPipeline.

BackendConfig

Runtime options for the compiled whisper.cpp backend.

LinearResampler

Streaming linear-interpolation resampler.

OnlineAsrConfig

Tunable parameters for a streaming ASR processor.

OnlineAsrModel

Loaded Whisper model that can create multiple streaming processors without reloading model weights from disk.

OnlineAsrProcessor

Streaming wrapper around whisper-rs implementing LocalAgreement-2.

ProcessOutput

Result of one streaming processing pass.

VadConfig

Tunable parameters for the integrated Silero VAD. Default matches whisper.cpp’s recommended Silero settings (250 ms minimum speech, 100 ms minimum silence, 0.5 probability threshold) plus a 2 s silence-to-reset window suitable for live microphone use.

VadModel

Loaded Silero VAD model that can be shared by multiple processors.

Word

One recognised word with absolute (offset-applied) timestamps in seconds.

Enums

DecodingStrategy

Whisper decoding strategy used for each streaming pass.

Error

All failures surfaced by yamabiko_whisper.

Constants

SAMPLE_RATE

Audio sample rate expected by OnlineAsrProcessor::insert_audio_chunk.

Traits

AudioSample

Primitive audio sample types that can be normalized to f32 PCM.

Functions

downmix_interleaved

Downmix interleaved source audio to mono f32 PCM.

install_log_hooks

Forward whisper.cpp / GGML / VAD logs to a log / tracing backend. Without calling this they are silently dropped, which is usually what you want; call this once at startup if you do want to see them. Thin wrapper around whisper_rs::install_logging_hooks so callers do not need to depend on whisper-rs directly.