OpenAI Whisper replaces Otter.ai as a self-hosted, free transcription option

At a glance:

• OpenAI Whisper can transcribe and translate audio locally with no cloud costs, supporting roughly 99 languages and multiple task types in a single model.
• The open-source tool runs on consumer hardware (including MacBooks) with model sizes from tens of millions to about 1.5 billion parameters, trading speed for accuracy.
• Unlike Otter and Fireflies, Whisper does not join meetings, send emails, or process data on external servers, removing the privacy and consent concerns some users flag.

Why users are moving away from Otter and Fireflies

I think dictation or transcription is one of the best use cases of AI. Take tools like Wispr Flow, for example. I’ve been using it a lot recently, and it has made me much faster at typing than I used to be. Similarly, there are tools like Otter, which can transcribe meetings, whether they’re virtual or in person. That said, I am not exactly a fan of Otter and similar tools like Fireflies because they are intrusive. Once you sign up and let them join a meeting, they send emails to everyone on the call without taking any consent. While the transcription quality is fairly accurate and the features are sometimes useful, it’s hard to ignore the privacy concerns. If a company is willing to bypass consent and send emails on your behalf, it raises questions about how it handles the data it collects.

This pattern has pushed some users to look for alternatives that keep recordings and processing on their own machines. The calculus is not only about accuracy or price; it is also about who gets to see, store, or forward sensitive conversations. When a service can invite itself into a call and message participants without explicit permission, it reframes the risk profile for teams handling confidential or regulated information. Even with encryption or compliance assurances, trusting infrastructure you cannot inspect remains a sticking point for many.

How Whisper works and what it was trained on

Whisper was trained on about 680,000 hours of multilingual audio collected from the internet, which is significantly larger than most speech recognition datasets. I don't have a typical English accent, and it still handles my voice quite well without mispronouncing words. It even handles background noise and messy real-world audio well. The best part about Whisper is that it’s open-source under the MIT license, and the full model weights are publicly available. You can run it entirely on your own machine without relying on a cloud service. You just need to point it to a local audio file, and it returns a transcript. Otter processes audio on its servers, which means your recordings leave your system.

Whisper focuses on transcription without extending into your workflow. It does not join meetings, send emails, or interact with other participants. It processes audio and outputs text, and that’s it. The model supports transcription, English translation, language detection, and timestamp prediction within a single system. It works across roughly 99 languages, though accuracy depends on the amount of training data available for each language. I've not had much success with languages other than English, Spanish, and French. Under the hood, Whisper uses a Transformer-based encoder-decoder architecture that converts audio into text tokens. It can handle multiple tasks, such as speech recognition and translation, without separate pipelines.

Choosing the right model size for your hardware

There are multiple model sizes available, ranging from small models with tens of millions of parameters to large models with around 1.5 billion parameters. I tested the smaller models, and while they are fast, they are not very accurate with the transcription. These are something you can use if your server or your PC is not very powerful. But even if you have a decent MacBook, you should consider running a model with at least a billion parameters. It will offer a lot better accuracy and speed.

This choice matters because it directly affects both usability and privacy trade-offs. Smaller models can run on modest CPUs or low-power devices, making Whisper accessible in offline or air-gapped environments. Larger models demand more memory and compute but narrow the gap with cloud services on transcription quality. For teams that cannot risk uploading board meetings or customer calls, accepting slower or heavier local inference is often preferable to exposing data to third-party servers.

Setting up Whisper locally on a PC

Getting Whisper running on your machine takes a few steps, but it is fairly easy once you know what to install. Whisper is distributed as a Python library, so the first requirement is a working Python setup. You need Python and pip installed on your system. If you already use Python for anything else, you can move straight to installing Whisper. Otherwise, installing Python from the official website is enough, since pip comes bundled with it in most cases.

Once Python is ready, you install Whisper directly from its GitHub repository: pip install git+https://github.com/openai/whisper.git. This pulls the latest version of the model and its dependencies. The install process also brings in PyTorch, which Whisper uses under the hood for running inference. The second requirement is FFmpeg, which handles audio and video decoding. Once both Whisper and FFmpeg are installed, you are ready to transcribe files.

Running your first transcription and supported formats

Move into the folder that contains your audio or video file and run: whisper --model base --language en --task transcribe your_audio_file.mp3. You can replace the filename with any supported format. Whisper supports WAV, MP3, M4A, FLAC, and even video formats like MP4 and MKV, since FFmpeg handles the conversion internally. You do not need to extract audio beforehand.

When you run the command, Whisper processes the file locally and shows progress in the terminal. After it finishes, it saves the transcript to your machine. By default, you get a plain text file, but it can also generate subtitle formats like SRT if needed. For users who want control over cost, privacy, and workflow, this setup provides a concrete alternative to subscription-based services, with the added benefit of offline operation and no recurring fees.