LOKAL is a user-friendly application to do AI transcriptions locally (i.e. on your own computer).
- Low cost. Payment is voluntary, and very small.
- High quality. Quality is on-par with on-cloud alternatives.
- Privacy-aware. LOKAL avoids the need to send audio data to third parties. It is also possible to run transcriptions entirely offline.
- Open-source. Commercial use is allowed under the license terms.
Interviews, focus groups, statements, and other oral stories can hold biometric data, personal identifiable information (PII), and sensitive details. The risks of sharing such data with untrusted 3rd parties is significant, including a potential for audio "deepfakes".
LOKAL runs locally on your computer.1 It can avoid you the need to send audios out.
If you want to save time, use the standalone version of LOKAL.
- Download the executable (.exe) installation file available here.
- Double-click.
- Install.
- Use.
You can also use LOKAL as a pure Python app. Installation and usage are much harder, but you get more freedom and can tinker with the code. See instructions at the end of this README.
While LOKAL can work without it, some features require FFmpeg – installation guidance is available here.
- Audio conversion. Audios in .wav format work even without FFmpeg. That said, FFmpeg is required to transcribe audios in formats other than .wav.
- OpenAI's Whisper. Systran's 'Faster Whisper' work without FFmpeg. That said, FFmpeg is required to use the original OpenAI's 'Whisper' models.
LOKAL offers users a number of choices. It's a long list, but once you open the app you'll see thing are fairly intuitive.
- Select audio. Press select audio to choose the audio to transcribe.
- Currently, only .wav files are supported. If you do not see your audio, make sure it's in .wav format.
- Terms and conditions. Use the checkbox to agree to the terms and conditions of usage.
- Click on the link to read terms and conditions.
- Agree to terms and conditions using the checkbox.
- Model (family). Use the dropwodn to choose the family of models to use for transcription.
- Currently, LOKAL offers two model families, OpenAI's Whisper and Systran's take on Whisper, called "faster-whisper".
- OpenAI's Whisper requires FFmpeg. Systran's Faster Whisper does not.
- The goal is to add more options.
- Model (size). Use the dropdown to specify the model size to use for transcription.
- Models come in different sizes. The larger the model, the better the output. The larger the model, the longer the transcription takes.
- Rule of thumb. If you can understand the conversation easily, one of the smaller models might be enough.
- Approach. Use the dropdown to choose between three different approaches to transcription, meant for different types of conversations.
- Simple. A straightforward word-by-word transcription. Best for complex/dynamic conversations with many speakers and interruptions, where it is very difficult to track of all speaker changes.
- Segmentation. A transcription split into paragraphs roughly following pauses and changes of speakers. Best for semi-structured conversations like interviews and focus groups, where tracking speaker changes is easier.
- Diarisation. A transcription split into blocks roughly corresponding to segments by specific speakers (with speaker labels). Best for structured conversations where speakers have clear turns and speaker changes are clear, like panels or seminars.
- Language. Use the dropdown to select the language for transcription.
- If left blank, the model tries to identify the language by itself.
- Specifying the language might make it easier for the model and speed things a little.
Optional hyper-parameters appear and dissapear depending on user selections. All such fields have a preset value. There is no need to even touch them. That said, users can also adjust them at will.
- Ignore short segments.
- Appears when users choose 'segmentation'.
- Use to set a threshold to ignore short words/comments like fillers or interruptions.
- If treshold is too high, it might ignore parts of sentences before/after pauses.
- Ignore short pauses.
- Appears when users choose 'segmentation' or 'diarisation'.
- Use to set a threshold to ignore short pauses that might cause excessive paragraph breaks in the final transcript.
- If this hyper-parameter is set too high, the models might ignore legitimate changes between speakers.
- Number of speakers.
- Appears when users choose 'diarisation'.
- Enables users to enter the exact number of speakers.
- There is no need to set a number. The diarisation model can try and guess the number of speakers. That said, results can be better if the model knows the number of speakers in advance.
LOKAL is not a guarantee of privacy. That said, LOKAL can help you manage privacy in the following ways:
- LOKAL avoids you the need to send audios to third-party service providers. This can help reduce risk exposure, especially if dealing with sensitive or easily identifiable audios.
- LOKAL allows you to run transcriptions offline. Running a few mock transcriptions while online is needed to trigger the download of models. After your chosen models have downloaded, however, you can disconnect from the Internet and everything should works just fine.
- If you need or want maximal privacy, you can also copy LOKAL's installation (LOKAL installs to '~/<-username->/AppData/Local/Programs/LOKAL') to an island computer after triggering the download of models.
LOKAL aims to assist humans with preparatory tasks rather than replace humans. Reducing the costs and time of initial speech-to-text conversion allows users to assign more resources on editing.
For the same reason, users remain in control and are responsible for the final quality of transcriptions. Users should therefore keep in mind all AI models and systems have limitations.
LOKAL's limitations include but are not limited to:
- All models may present potentially high error rates.
- The segmentation and diarisation approaches and models used in these approaches can struggle with overlapping speakers.
- The transcription models can struggle with names, places, acronyms, and potentially, accents. Users must not use outputs without editing them.
- The transcription models can struggle with conversations taking place in different languages – an idea is to do multiple runs with different language selections and then join the results.
We made LOKAL as easy to install and use as possible. However do get in touch if you feel stuck: hello@polyzentrik.com.
LOKAL (transcriptions) is a product by polyzentrik.com, released under an Apache 2.0 open source license. The code for LOKAL is available via GitHub.
LOKAL implements a voluntary activation approach with a soft free trial:
- Core functionality works regardless of whether LOKAL is activated
- Activation enables visual styles
- Activation is also needed to keep the user interface pretty after five free trial transcriptions
Activation is possible using the invoice number you get after making a very small voluntary payment.
.
.
.
Skip this section if not a coder. It is REALLY hard.
You can also use LOKAL as a pure Python app.
Assuming you have Python (3.11.0) installed on your computer, you need to:
- Clone this repository.
- The main branch has the latest released version.
- The dev branch has latest commits, to be shipped on next release.
- For guidance, see https://docs.github.com/en/repositories/creating-and-managing-repositories/cloning-a-repository and https://stackoverflow.com/questions/1778088/how-do-i-clone-a-single-branch-in-git.
- Create an environment around it.
- For guidance, see https://docs.python.org/3/library/venv.html.
- Install required libraries using pip install -r requirements.txt.
- For guidance, see https://pip.pypa.io/en/stable/user_guide/.
- Check this file for some tiny but necessary library adjustments needed for progress bars to render.
- Run LOKAL using python lokal.py or python -m lokal.
- For guidance, see https://pythonbasics.org/execute-python-scripts/.
.
.
.
Footnotes
Footnotes
-
An Internet connection is needed for initial model download. The first time you choose a model, the model downloads. After, it is possible to perform transcriptions offline. Resetting models is possible from within LOKAL, which deletes all models in memory and forces a re-download next time a model is used. ↩