About Easy Whisper

The python library easy_whisper is an easy to use adaptation of the popular OpenAI Whisper for transcribing audio files. The main features are:

• both CLI and (tkinter) GUI user interface
• faster processing of long audio even on CPU
• output in .txt format with time stamps in milliseconds precision

Installation

In the terminal, run:

pip install easy_whisper

OpenAI Whisper requires FFmpeg to process non-WAV files. A brief installation guide can be found in the Github repository of OpenAI Whisper:

Use GUI

Simply type in the command line:

easy_whisper

or:

python -m easy_whisper

to start the tkinter window.

Mouseover a control element to show quick help info.

When used for the first time, it can take longer, as it has to be checked if CUDA (faster) is available or only CPU (slower). Also the modules have to be compiled for faster use later. Moreover, OpenAI Whisper models which have not yet been used must first be downloaded. For more info about models, read the section Model below.

In the following, I'll briefly explain the control elements one by one.

Path

On the first run, the path is set to the most recent audio file in the current working directory, which, of course, can be changed by clicking on the "Choose File" button on the right.

If the preselected "save setting" option is not unselected, then on the next start, the folder used last time will be automatically chosen. This is practical if the user adds audio files to the destined folder one at a time to transcribe them immediately after adding.

Language

Specify the language of the audio file can enhance the speed. If the language is unknown, you can also choose Auto for auto detection. Auto is preselected on the first run. Later, if "save setting" is not unselected, then the language used last time will be preselected.

Model

For higher speed, choose a smaller OpenAI Whisper model, for higher quality, chose a larger one. The models with the suffix "en" are for English only. I recommend "base.en" for English and "small" for other languages, as the quality is already high enough for most purposes.

To learn more about the models, view OpenAI Whisper:

in tmp

By default, easy_whisper splits the audio file into sentences and stores the interim files in a folder named tmp. If the source audio file has the path folder1/file1.mp3, then the tmp folder has the path folder1/tmp.

Sometimes, you may want to split the file manually. One example is that you have recorded the file and processed it with Audacity to trim the start and end part, then it is convenient to split the file subsequently there. For this purpose, go in Audacity to "Analyze/Stille-Finder", then click "OK", this way you can see immediately how the file is split. If you are satisfied with the result, go to "File/Export/Export Multiple", select "Include audio before first label" and change "First file name" to "S", set file format to WAV, then choose the tmp folder you have created and click "Export".

After this is done, you can select the option "in tmp" in the tkinter window so that easy_whisper knows it doesn't have to split the file again. If you forget to do so, easy_whisper still works, the only drawback is it will unnecessarily spend time to split the file.

Audacity also uses FFmpeg to process non-standard audio files. However, the FFmpeg you installed for OpenAI Whisper possibly doesn't work for Audacity. If so, you can download the proper FFmpeg from Audacity Support:

del tmp

By default, the tmp folder with its content remains after transcription is done. If any interim audio file causes error during the transcription, a report .txt file will be placed next to the audio file for inspection. For example, if the interim file 02-S.wav causes error, then a report file 02-S_report.txt will be placed in the same tmp folder. If you see no need for keeping the tmp folder, you can check the "del tmp" option and the folder will be automatically deleted after the transcription job is done.

short

easy_whisper is rather for transcribing audiobooks ranging from few minutes to more than an hour. To transcribe short dialogues with one sentence you don't need this library. But of course you can still use easy_whisper if you want. In this case, you can check this option to tell easy_whisper not to split the audio in sentences. It still works if you leave this option unchecked, the only drawback is easy_whisper will unnecessarily spend additional time.

MSL

easy_whisper enhances speed for transcribing longer audio by splitting the speech into sentences, which are detected at the pause between them. The pause, or MSL (minimum silence length), can be adjusted here. It is recommended to leave the default of 900ms unchanged, but for speakers tending to pause significantly longer or shorter between sentences/clauses you can also adjust this parameter up or down, respectively.

STH

The silence is detected at the specified threshold, or STH (silence threshold). The default value is -16dB and it is recommended to leave it unchanged.

save setting

Preselected by default, only unselect if necessary. If selected, settings for Path/Language/Model/task(transcribe or translate)/MSL/STH/in_tmp/del_tmp from last usage will be saved for the use next time. It is practical to leave this option selected if the user intends to transcribe files with similar settings over a longer time period. The default models for English and other Languages (including Auto) are stored separately - one for English only and one for other languages.

Reset

Click this button if you want to reset Path/Language/Model/task(transcribe or translate)/MSL/STH/in_tmp/del_tmp to the system default values, which are current folder/Auto (detection)/small or base.en/transcribe/900ms/-16dB/no/no

transcribe/translate

By default, easy_whisper transcribes the audio file, but you can also switch to "translate", in which case the (non-English) speech will be translated into English. Click the "Run" button to start transcription/translation, the window is then minimised and you can do something else while waiting. When the job is done, the window reappears to regain your attention. The result is a .txt file with the same name and put in the same folder as the audio file. For example, if the path to the audio file is folder2/audio2.wav, then the path to the .txt file will be folder2/audio2.txt. In case of translation, the suffix "_English" will be added, so the output path is then folder2/audio2_English.txt. The .txt file contains the transcription/translation with time stamps, which looks like below:

[00:00:00,000 --> 00:00:04,000] Chapter 1. Title. [00:00:04,140 --> 00:00:08,900] Sentence one. ... ...

Use CLI

To use the CLI (command line interface), simply write "easy_whisper" followed by any argument. For example:

easy_whisper folder3/audio3.mp4

transcribes audio3.mp4 or translates it if the setting "translation" was used and saved last time. If you want to use all saved settings including the file path, type:

easy_whisper --cli

or:

easy_whisper -c

then the most recent audio file in the folder accessed last time will be processed.

The arguments largely correspond to the tkinter GUI control elements, with the addition of the --cli argument. Below is a summary table:

Use --help or -h to see help information.

Disclaimer

I wrote easy_whisper for my personal use and published it for others who may also find it useful. If you have any question, feel free to ask, but keep in mind that I can only reply in my spare time.