Crate transcribe_rs 

transcribe-rs

A Rust library providing unified transcription capabilities using multiple speech recognition engines.

Features

• ONNX Models: SenseVoice, GigaAM, Parakeet, Moonshine (requires onnx feature)
• Whisper: OpenAI Whisper via GGML (requires whisper-cpp feature)
• Whisperfile: Mozilla Whisperfile server (requires whisperfile feature)
• Remote: OpenAI API (requires openai feature)
• Timestamped Results: Detailed timing information for transcribed segments
• Unified API: SpeechModel trait for all local engines
• Hardware Acceleration: GPU support for ORT engines (ort-cuda, ort-rocm, ort-directml, ort-coreml, ort-webgpu) and whisper.cpp (Metal/Vulkan) via the accel module

Backend Categories

This crate provides two categories of transcription backend:

• Local models implement SpeechModel and run inference in-process or via a local binary. This includes all ONNX models, Whisper (via whisper.cpp), and Whisperfile.
• Remote services implement [RemoteTranscriptionEngine] (requires openai feature) and make async network calls to external APIs. This includes OpenAI.

These traits are intentionally separate because the execution model differs: local models are synchronous and take audio samples directly, while remote services are async and may only accept file uploads.

Quick Start

[dependencies]
transcribe-rs = { version = "0.3", features = ["onnx"] }

use std::path::PathBuf;
use transcribe_rs::onnx::sense_voice::{SenseVoiceModel, SenseVoiceParams};
use transcribe_rs::onnx::Quantization;
use transcribe_rs::SpeechModel;

let mut model = SenseVoiceModel::load(
    &PathBuf::from("models/sense-voice"),
    &Quantization::Int8,
)?;

let result = model.transcribe(&samples, &transcribe_rs::TranscribeOptions::default())?;
println!("Transcription: {}", result.text);

Audio Requirements

Input audio files must be:

• WAV format
• 16 kHz sample rate
• 16-bit samples
• Mono (single channel)

Migrating from 0.2.x to 0.3.0

Version 0.3.0 is a breaking release. If you need the old API, pin to version = "=0.2.9".

SpeechModel::transcribe signature changed:

// Before (0.2.x):
model.transcribe(&samples, Some("en"))?;
model.transcribe_file(&path, None)?;

// After (0.3.0):
use transcribe_rs::TranscribeOptions;
model.transcribe(&samples, &TranscribeOptions { language: Some("en".into()), ..Default::default() })?;
model.transcribe_file(&path, &TranscribeOptions::default())?;

SpeechModel now requires Send, enabling Box<dyn SpeechModel + Send> for use across threads.

TranscribeOptions includes a translate field (default false). Engines that support translation (Whisper, Whisperfile) will translate to English when set to true.

Whisper capabilities are now dynamic. WhisperEngine::capabilities() returns the actual language support of the loaded model (English-only vs multilingual) rather than always reporting all 99 languages.

Re-exports

pub use accel::get_ort_accelerator;

pub use accel::get_whisper_accelerator;

pub use accel::get_whisper_gpu_device;

pub use accel::set_ort_accelerator;

pub use accel::set_whisper_accelerator;

pub use accel::set_whisper_gpu_device;

pub use accel::OrtAccelerator;

pub use accel::WhisperAccelerator;

pub use accel::GPU_DEVICE_AUTO;

pub use error::TranscribeError;

Modules

accel

Per-engine accelerator preferences.

audio

Audio processing utilities for transcription.

error

transcriber

Chunked transcription strategies.

vad

Voice Activity Detection (VAD).

Structs

ModelCapabilities

Describes the capabilities of a speech model.

TranscribeOptions

Options for transcription.

TranscriptionResult

The result of a transcription operation.

TranscriptionSegment

A single transcribed segment with timing information.

Traits

SpeechModel

Unified interface for speech-to-text models.