Crate whisper_apr 

Whisper.apr

WASM-first automatic speech recognition engine implementing OpenAI’s Whisper architecture.

Overview

Whisper.apr is designed from inception for WASM deployment via wasm32-unknown-unknown, leveraging Rust’s superior WASM toolchain for:

• 30-40% smaller binary sizes through tree-shaking
• Native WASM SIMD 128-bit intrinsics without Emscripten overhead
• Zero-copy audio buffer handling via shared memory

Quick Start

use whisper_apr::{WhisperApr, TranscribeOptions};

let whisper = WhisperApr::load("base.apr")?;
let result = whisper.transcribe(&audio_samples, TranscribeOptions::default())?;
println!("{}", result.text);

Features

• std (default): Standard library support
• wasm: WASM bindings via wasm-bindgen
• simd: SIMD acceleration via trueno
• tracing: Performance tracing via renacer

Re-exports

pub use error::WhisperError;

pub use error::WhisperResult;

pub use error::WhisperError;

pub use error::WhisperResult;

Modules

audio

Audio preprocessing module

backend

Backend abstraction and automatic selection Backend abstraction and selection (WAPR-140 to WAPR-141)

benchmark_generated

Benchmark infrastructure for multi-backend comparison Benchmark infrastructure for multi-backend performance comparison

detection

Language detection module

diarization

Speaker diarization module (who spoke when) Speaker diarization module (WAPR-150 to WAPR-153)

error

Error types for Whisper.apr

format

APR Model Format (v2)

gpu

WebGPU compute backend for accelerated inference WebGPU compute backend for accelerated inference (WAPR-120 to WAPR-143)

inference

Inference engine

memory

Memory management for Whisper inference

model

Model loading and inference

parallel

Unified parallelism abstraction for CLI and WASM (§11.3.2) Unified parallelism abstraction for CLI and WASM

probe

Activation probing for forward-pass debugging (WAPR-MOONSHINE-013) Activation probing infrastructure for forward-pass debugging

progress

Progress tracking and callbacks

publish

HuggingFace Hub publishing (WAPR-PUB-001) HuggingFace Hub Publishing Module (WAPR-PUB-001)

realizar_inference

Re-exports world-class production inference primitives from realizar.

simd

SIMD-accelerated operations via trueno

timestamps

Timestamp extraction and word-level alignment (WAPR-160 to WAPR-163)

tokenizer

BPE tokenizer

trace

Tracing utilities for pipeline instrumentation

vad

Voice Activity Detection (VAD)

verify

Pre-publish verification (WAPR-PUB-001) Pre-publish Verification Module (WAPR-PUB-001)

vocabulary

Vocabulary and hotword customization Custom vocabulary fine-tuning module (WAPR-170 to WAPR-173)

Macros

trace_enter

Enter a tracing span (no-op when tracing feature is disabled)

trace_event

Log a tracing event (no-op when tracing feature is disabled)

trace_span

Create a tracing span (no-op when tracing feature is disabled)

Structs

BatchTranscriptionResult

Result of batch transcription (WAPR-083)

PartialTranscriptionResult

Result of partial transcription during streaming (WAPR-101)

ProfilingStats

Performance profiling statistics (WAPR-PERF-004)

Segment

A timestamped segment of transcription

StreamingSession

Streaming transcription session (WAPR-101)

SummarizeOptions

Options for LFM2 summarization (Phase 3 API - Section 18.5)

TranscribeOptions

Options for transcription

TranscribeSummaryResult

Result of transcription with summarization (Phase 3 API - Section 18.5)

TranscriptionResult

Result of transcription

VadSpeechSegment

A speech segment detected by VAD (WAPR-093)

VadTranscriptionResult

Result of VAD-triggered transcription (WAPR-093)

WhisperApr

Main Whisper ASR engine

Enums

DecodingStrategy

Decoding strategy for transcription

ModelType

Whisper model configuration

Task

Task type for transcription