Crate scriptrs 

Rust transcription with native CoreML Parakeet v2 inference

scriptrs is currently a narrow macOS-first transcription crate:

• macOS only
• CoreML only
• Parakeet TDT v2 only
• mono 16kHz &[f32] input

The base crate exposes a single-chunk TranscriptionPipeline. Fast long-audio chunking and overlap merging are available behind the long-form feature via [LongFormTranscriptionPipeline]. VAD-backed speech region planning is an additional long-form-vad feature.

Choosing a pipeline

Use TranscriptionPipeline when your audio already fits in one Parakeet window. If the input is too long, it returns TranscriptionError::AudioTooLong.

Use [LongFormTranscriptionPipeline] with the long-form feature when you want scriptrs to own long-audio chunking internally. This default path is tuned for speed and works well on dense, mostly continuous speech.

Add long-form-vad when you need VAD-backed speech region planning for sparse speech, long silences, or recordings with a lot of non-speech audio.

Model loading

With the default online feature, TranscriptionPipeline::from_pretrained and [LongFormTranscriptionPipeline::from_pretrained] download the runtime bundle from avencera/scriptrs-models on Hugging Face.

If you already manage models yourself, use TranscriptionPipeline::from_dir or [LongFormTranscriptionPipeline::from_dir] with a local bundle layout:

models/
  parakeet-v2/
    encoder.mlmodelc/
    decoder.mlmodelc/
    joint-decision.mlmodelc/
    vocab.txt

With long-form-vad, add:

models/
  vad/
    silero-vad.mlmodelc/

You can also override model resolution with environment variables:

• SCRIPTRS_MODELS_DIR=/path/to/models
• SCRIPTRS_MODELS_REPO=owner/repo

Input requirements

scriptrs expects mono 16kHz audio samples as &[f32]. File decoding stays outside the core library on purpose.

Examples

Single-chunk transcription:

use scriptrs::TranscriptionPipeline;

let audio = load_audio();
let pipeline = TranscriptionPipeline::from_pretrained()?;
let result = pipeline.run(&audio)?;

println!("{}", result.text);

Long-form transcription:

use scriptrs::LongFormTranscriptionPipeline;

let audio = load_audio();
let pipeline = LongFormTranscriptionPipeline::from_pretrained()?;
let result = pipeline.run(&audio)?;

println!("{}", result.text);

VAD-backed long-form transcription:

use scriptrs::{LongFormConfig, LongFormMode, LongFormTranscriptionPipeline};

let audio = load_audio();
let pipeline = LongFormTranscriptionPipeline::from_pretrained()?;
let config = LongFormConfig {
    mode: LongFormMode::Vad,
    ..LongFormConfig::default()
};
let result = pipeline.run_with_config(&audio, &config)?;

println!("{}", result.text);

Structs

ModelBundle

Resolved model paths for scriptrs

ModelManager

Downloads and caches model bundles from Hugging Face

TimedToken

A single token with timing metadata

TranscriptChunk

A text chunk in the final transcript

TranscriptionConfig

Configuration for single-chunk transcription

TranscriptionPipeline

Single-chunk Parakeet v2 transcription pipeline

TranscriptionResult

Final transcription output

Enums

TranscriptionError

Errors returned by scriptrs