Speech recognition tool to convert audio to text transcripts, for Linux and Raspberry Pi.

spchcat is a command-line tool that reads in audio from .WAV files, a microphone, or system audio inputs and converts any speech found into text. It runs locally on your machine, with no web API calls or network activity, and is open source. It is built on top of Coqui's speech to text library, TensorFlow, KenLM, and data from Mozilla's Common Voice project.

It supports multiple languages thanks to Coqui's library of models. The accuracy of the recognized text will vary widely depending on the language, since some have only small amounts of training data. You can help improve future models by contributing your voice.

On Debian-based x86 Linux systems like Ubuntu you should be able to install the latest .deb package by downloading and double-clicking it. Other distributions are currently unsupported. The tool requires PulseAudio, which is already present on most desktop systems, but can be installed manually.

There's a notebook you can run in Colab at notebooks/install.ipynb that shows all installation steps.

To install on a Raspberry Pi, download the latest .deb installer package and either double-click on it from the desktop, or run dpkg -i ~/Downloads/spchcat_0.0-2_armhf.deb from the terminal. It will take several minutes to unpack all the language files. This version has only been tested on the latest release of Raspbian, released October 30th 2021, and on a Raspberry Pi 4. It's expected to fail on Raspberry Pi 1's and 0's, due to their CPU architecture.

After installation, you should be able to run it with no arguments to start capturing audio from the default microphone source, with the results output to the terminal:

spchcat

After you've run the command, start speaking, and you should see the words you're saying appear. The speech recognition is still a work in progress, and the accuracy will depend a lot on the noise levels, your accent, and the complexity of the words, but hopefully you should see something close enough to be useful for simple note taking or other purposes.

If you don't have a microphone attached, or want to transcribe audio coming from another program, you can set the --source argument to 'system'. This will attempt to listen to the audio that your machine is playing, including any videos or songs, and transcribe any speech found.

spchcat --source=system

One of the most common audio file formats is WAV. If you don't have any to test with, you can download Coqui's test set to try this option out. If you need to convert files from another format like '.mp3', I recommend using FFMPeg. As with the other source options, spchcat will attempt to find any speech in the files and convert it into a transcript. You don't have to explicitly set the --source argument, as long as file names are present on the command line that will be the default.

spchcat audio/8455-210777-0068.wav 

If you're using the audio file from the test set, you should see output like the following:

TensorFlow: v2.3.0-14-g4bdd3955115
 Coqui STT: v1.1.0-0-gf3605e23
your power is sufficient i said 

You can also specify a folder instead of a single filename, and all .wav files within that directory will be transcribed.

So far this documentation has assumed you're using American English, but the tool will default to looking for the language your system has been configured to use. It first looks for the one specified in the LANG environment variable. If no model for that language is found, it will default back to 'en_US'. You can override this by setting the --language argument on the command line, for example:

spchcat --language=de_DE

This works independently of --source and other options, so you can transcribe microphone, system audio, or files in any of the supported languages. It should be noted that some languages have very small amounts of data and so their quality may suffer. If you don't care about country-specific variants, you can also just specify the language part of the code, for example --language=en. This will pick any model that supports the language, regardless of country. The same thing happens if a particular language and country pair isn't found, it will log a warning and fall back to any country that supports the language. For example, if 'en_GB' is specified but only 'en_US' is present, 'en_US' will be used.

All of these models have been collected by Coqui, and contributed by organizations like Inclusive Technology for Marginalized Languages or individuals. All are using the conventions for Coqui's STT library, so custom models could potentially be used, but training and deployment of those is outside the scope of this document. The models themselves are provided under a variety of open source licenses, which can be inspected in their source folders (typically inside /etc/spchcat/models/).

By default spchcat writes any recognized text to the terminal, but it's designed to behave like a normal Unix command-line tool, so it can also be written to a file using indirection like this:

spchcat audio/8455-210777-0068.wav > /tmp/transcript.txt

If you then run cat /tmp/transcript.txt (or open it in an editor) you should see `your power is sufficient i said'. You can also pipe the output to another command. Unfortunately you can't pipe audio into the tool from another executable, since pipes aren't designed for non-text data.

There is one subtle difference between writing to a file and to the terminal. The transcription itself can take some time to settle into a final form, especially when waiting for long words to finish, so when it's being run live in a terminal you'll often see the last couple of words change. This isn't useful when writing to a file, so instead the output is finalized before it's written. This can introduce a small delay when writing live microphone or system audio input.

It's possible to build all dependencies from source, but I recommending downloading binary versions of Coqui's STT, TensorFlow Lite, and KenLM libraries from github.com/coqui-ai/STT/releases/download/v1.1.0/native_client.tflite.Linux.tar.xz. Extract this to a folder, and then from inside a folder containing this repo run to build the spchcat tool itself:

make spchcat LINK_PATH_STT=-L../STT_download

You should replace ../STT_download with the path to the Coqui library folder. After this you should see a spchcat executable binary in the repo folder. Because it relies on shared libraries, you'll need to specify a path to these too using LD_LIBRARY_PATH unless you have copies in system folders.

LD_LIBRARY_PATH=../STT_download ./spchcat

The previous step only built the executable binary itself, but for the complete tool you also need data files for each language. If you have the gh GitHub command line tool you can run the download_models.py script to fetch Coqui's releases into the build/models folder in your local repo. You can then run your locally-built tool against these models using the --languages_dir option:

LD_LIBRARY_PATH=../STT_download ./spchcat --languages_dir=build/models/

After you have the tool built and the model data downloaded, create_deb_package.sh will attempt to package them into a Debian installer archive. It will take several minutes to run, and the result ends up in spchcat_0.0-2_amd64.deb.

There's a notebook at notebooks/build.pynb that runs through all the build steps needed to downloaded dependencies, data, build the executable, and create the final package. These steps are run inside an Ubuntu 18.04 Docker image to create the binaries that are released.

sudo docker run -it -v`pwd`:/spchcat ubuntu:bionic bash

Tool code written by Pete Warden, pete@petewarden.com, heavily based on Coqui's STT example. It's a pretty thin wrapper on top of Coqui's speech to text library, so the Coqui team should get credit for their amazing work. Also relies on TensorFlow, KenLM, data from Mozilla's Common Voice project, and all the contributors to Coqui's model zoo.

Tool code is licensed under the Mozilla Public License Version 2.0, see LICENSE in this folder.

All other libraries and model data are released under their own licenses, see the relevant folders for more details.