#transcription #microphone #speech-to-text #record #audio-stream #speech-detection

ds-transcriber

A crate using DeepSpeech bindings to convert mic audio from speech to text

5 releases (2 stable)

1.0.1 Mar 22, 2022
1.0.0 Mar 5, 2022
0.1.3 May 31, 2021

#769 in Audio

24 downloads per month

MIT license

18KB
239 lines

ds-transcriber

GitHub Crates.io docs.rs

A DeepSpeech powered utility for getting microphone stream transcriptions

Prelude

You can think of this crate as a wrapper for RustAudio's deepspeech-rs. It aims to provide transcription for microphone streams with optional denoising see cargo-features below.

Getting Started

This example shows the quickest way to get started with ds-transcriber. First, add ds-transcriber to your Cargo.toml

ds-transcriber = "1"

Download the DeepSpeech native client and then add its directory to your LD_LIBRARY_PATH and LIBRARY_PATH variables.

Have a look at StreamSettings to fine tune the transcription stream to parameters that better suit your environment

let mut model = ds_transcriber::model::instance_model(
    "model_file.pbmm",
    Some("scorer_file.scorer"),
)?;
let config = ds_transcriber::StreamSettings::default();
let i_said = ds_transcriber::transcribe(config, &mut model)?;
println!("I said: {}", i_said);

Rinse and repeat the last two lines

Cargo Features

This crate provides an optional feature of denoising of the audio stream (may result in better transcription). It is disabled by default, to enable it: use either the denoise or full key in the crate's features list.

ds-transcriber = { version = "1", features = ["denoise"] } # or features = ["full"]

Extras

This crate contains an example to get you started. Clone the repository and run it:

For help with arguments, run:

cargo run --example transcribe -- -h

To start the example, run

cargo run --example transcribe -- -m model_path -c deepspeech_native_client_dir

An optional (but recommended) argument for a language model (scorer) path can be provided with -s or --scorer

Re-exports

This crate also re-exports the deepspeech and nnnoiseless crates (if the denoise feature is enabled). You can use these re-exports instead of also depending on them separately.

Transcription Tips

Downloading the DeepSpeech model alone will give you results that are passable, at best, (depending on your accent), if you want to significantly improve them, you might also want to download a language model/scorer. It helps in cases like: I read a book last night vs I red a book last night. Simply put the scorer in the same directory as your model. The crate will automatically set it when you create your ds_transcriber::model::DeepSpeechModel

If you want to train your own model, for the best results, look into Mimic Recording Studio, it gives you prompts to read from and automatically prepares your audio files with their respective transcriptions for training which you can then use for fine tuning

Contributions

Always welcome! Open an issue or a PR if you have something in mind

Dependencies

~1–6MB
~113K SLoC