100% Private
No Signup
Free Forever
One of 64 free AI tools by Mahmoud Zalt.
Free Voice Notes
A free, private voice notes tool that turns spoken audio into a written transcript and a short AI summary without sending a single byte to any server. It chains two open-source models that run entirely on-device through Hugging Face Transformers.js: OpenAI Whisper for speech recognition and DistilBART CNN for abstractive summarization. Record straight from your microphone or upload an existing audio file, and the tool first transcribes the speech with the whisper-base model, then condenses that transcript into a concise summary with the Xenova/distilbart-cnn-6-6 model, using the same pipeline() API as the Python transformers library. Both models execute on your own hardware via ONNX Runtime and WebAssembly, so your audio and the resulting text are never uploaded, logged, or stored. The models download once on first use and are cached by your browser, and they are the same files shared with the speech-to-text and text-summarizer tools, so if you have used those, voice notes start almost instantly.
Free and provided as is, without warranty. Use at your own risk. Terms
Turn audio into notes without sending it anywhere
Most voice transcription and meeting-notes services upload your audio to a server, run it through a hosted model, and send the text back, which means your recording leaves your control the moment you hit upload. This free voice notes tool takes the opposite approach: both the speech recognition model and the summarization model run inside your browser tab. Once the models are downloaded, every transcript and summary is generated on your own device, so the meeting, memo, or private recording you process never touches a server.
That privacy model makes it a strong fit for sensitive material such as internal meetings, legal discussions, customer calls, or unpublished interviews, where sending audio to a cloud transcription service is not acceptable. You get a full transcript plus a concise summary with the same privacy guarantees as an offline app, and you can verify there is no upload by watching the Network tab in your browser DevTools while you work.
OpenAI Whisper and DistilBART, chained in your browser
This tool is built on Transformers.js, the open-source library from Hugging Face whose goal is state-of-the-art machine learning for the web: it lets you run Transformers directly in your browser with no need for a server. It is designed to be functionally equivalent to the Hugging Face transformers Python library, exposing the same pipeline() API, and it executes models with ONNX Runtime compiled to WebAssembly so they run on ordinary CPUs without any backend.
Voice notes chain two models. First, the onnx-community/whisper-base build of OpenAI Whisper transcribes your audio into text, processing it in 30 second chunks with a short stride so nothing is lost at the boundaries. Then the Xenova/distilbart-cnn-6-6 model, a distilled BART fine-tuned on the CNN/DailyMail dataset for abstractive summarization, condenses that transcript into a short summary. The Whisper weights are roughly 80MB and the DistilBART weights roughly 240MB; both download once, are cached by your browser, and are shared with the speech-to-text and text-summarizer tools. Transformers.js is released under the permissive Apache 2.0 license.
Tips for accurate transcripts and useful summaries
Transcription quality follows audio quality. Record in a quiet space, keep the microphone close, and avoid people talking over each other, since background noise, crosstalk, and heavy accents all make speech recognition harder. If you are uploading a file, a clean recording at a normal speaking pace will transcribe far better than a noisy or distant one. The transcript is shown in full so you can read and correct it, and you can copy or download it as a text file independently of the summary.
The summary is abstractive, meaning DistilBART writes new sentences that paraphrase the meaning rather than copying lines verbatim. That reads more naturally and compresses ideas across the whole transcript, but because the model writes in its own words it can occasionally shift a fact or figure, so treat the summary as a fast first pass and verify critical names, numbers, and decisions against the transcript and the original audio. For very long recordings, the transcript stays complete while the summary covers as much as fits the model context window; split long sessions into sections for the best per-section summaries.
More Free Tools
More than 20 free AI tools.
The specific workflow this tool is built for
A meeting attendee records a call, gets both the full transcript for reference and a short summary they can paste directly into a team channel or a project management ticket, without spending time writing meeting notes by hand afterward. Someone dictating a quick voice memo while walking or driving turns it into structured text and a one-paragraph summary in the time it takes to review it once, rather than transcribing it manually or listening back later to remember what they said.
Students recording a lecture get a searchable transcript for looking up a specific point later, plus a condensed summary that works as a fast review before an exam without re-listening to the entire recording, and sales or support teams reviewing a call recording get both an exact record of what was said and a quick summary to scan before a follow-up conversation.
How this compares to Otter.ai and Fireflies.ai
Otter.ai and Fireflies.ai both offer automatic transcription plus AI-generated meeting summaries as their core paid feature, with plans typically running 10 to 30 dollars a month per user, and both work by joining or recording your meetings and processing the audio on their servers. That server-side processing is exactly how they deliver features like automatically joining a Zoom call, but it also means every meeting recorded through them lives on that company's infrastructure.
This tool covers the same core value, transcript plus summary, for free with the audio never leaving your device, at the cost of not having a bot that joins your calls automatically; you record or upload the audio yourself. For anyone who already has the recording in hand and just wants the transcript and summary without a subscription or a bot in the call, this covers the same ground for zero cost and stronger privacy.
Why two models chained together, and how to get useful action items
Running transcription and summarization as two separate specialized models, rather than one model trying to do both end-to-end, is a deliberate design choice: Whisper is trained specifically to convert audio into accurate text, and DistilBART is trained specifically to condense text into a summary, each does its one job better than a single generalist model would do both. This pipeline approach, chain focused models rather than one model that tries to do everything, is a common and effective pattern in practical machine learning systems.
To get a summary that actually reads like useful action items rather than a vague recap, it helps to speak decisions and next steps explicitly during the recording, saying "so the action item is Sarah will send the contract by Friday" gives the summarization model a clear, quotable sentence to work with, while a decision that is only implied across several minutes of discussion is much harder for any summarizer, human or AI, to extract cleanly.
How It Works
Click Start Recording to capture audio from your microphone, or click Upload Audio File to bring in an existing MP3, WAV, M4A, WebM, OGG, or FLAC recording.
On the first run the OpenAI Whisper speech model downloads into your browser and transcribes the audio, then the DistilBART summary model downloads and condenses the transcript.
Read the transcript and the AI summary in separate cards, then copy either one to your clipboard or download it as a text file.
Need expert help with AI?
Looking for a specialist to help integrate, optimize, or consult on AI systems? Book a one-on-one technical consultation with an experienced AI consultant to get tailored advice.
Key Features
Privacy & Trust
Use Cases
Limitations
- The first run downloads two models, roughly 80MB for Whisper and 240MB for DistilBART, which can take a few minutes on slower connections, though both are cached by your browser afterward
- Transcription accuracy depends on audio quality: background noise, crosstalk, heavy accents, and low microphone quality all reduce accuracy
- DistilBART has a fixed context window, so very long transcripts are truncated before summarization, and long recordings may need to be split into sections
- Abstractive summarization rephrases content in the model own words, so it can occasionally shift a fact or figure: verify important names, numbers, and claims against the transcript and the original audio
- Both models run on your CPU via WebAssembly, so processing is slower on phones and older machines than on a modern laptop or desktop
Frequently Asked Questions
Is this voice notes tool really free?
Yes, it is completely free with no signup, no account, and no usage limits. Because both the Whisper speech model and the DistilBART summary model run on your own hardware via Hugging Face Transformers.js instead of a paid cloud API, there are no per-minute or per-request costs to pass on. You can record, transcribe, and summarize as much audio as you want, as often as you want, without a credit card, an API key, or a rate limit.
Is my audio sent to a server when I record or upload it?
No. Both the transcription and the summarization run locally in your browser using Transformers.js and WebAssembly. After the two model files are downloaded once, all processing happens on your own device with zero network requests carrying your audio or text. Your recording, transcript, and summary are never uploaded, logged, stored, or seen by anyone, which makes this safe for confidential meetings, private memos, and sensitive recordings that should not touch a cloud service. You can confirm this for yourself by opening the Network tab in your browser DevTools while you process audio.
Which models does it use and how big are they?
It uses two models. For speech recognition it loads onnx-community/whisper-base, an ONNX build of OpenAI Whisper that is roughly 80MB. For summarization it loads Xenova/distilbart-cnn-6-6, a distilled BART fine-tuned on the CNN/DailyMail dataset that is roughly 240MB. Both are loaded through the Transformers.js pipeline() API, the in-browser equivalent of the Hugging Face transformers Python library. They download once on first use and are then cached by your browser so later runs start quickly.
Why is the first voice note slower than the ones after it?
The first run includes the one-time download of both models plus initialization of the ONNX Runtime and WebAssembly backend. After the weights are cached in your browser, later runs skip the downloads entirely and start much faster. The Whisper and DistilBART files are also shared with the speech-to-text and text-summarizer tools, so if you have already used those, the relevant model may already be cached. Speed still depends on your device CPU, since the models run locally on WebAssembly rather than on a remote GPU.
Can it summarize a long meeting or lecture?
It can transcribe long audio in 30 second chunks, so the transcript itself can be quite long. The summary step is the limiting factor: like all Transformer models, DistilBART has a fixed context window, so very long transcripts are truncated before summarization. For long recordings, the transcript is still complete and accurate, but for the best summary you may want to split a very long session into sections, summarize each part, and optionally summarize the combined summaries for a final overview.
Do I need to install anything to use it?
No installation is needed. It is a pure web tool that runs in any modern browser that supports WebAssembly and microphone access. The only things that download are the Whisper and DistilBART models themselves, which Transformers.js fetches from the Hugging Face Hub on first use and caches in your browser. There is no extension, no desktop app, and no Python or Node environment to set up.
Does it work offline and on mobile?
Once both models have been downloaded and cached, transcription and summarization run without an internet connection because everything happens on-device. Recording from the microphone and uploading files both work in mobile browsers that support WebAssembly, though phones and tablets are slower than a laptop or desktop and the first downloads use mobile data. For the smoothest experience, load the models once over Wi-Fi and then use the tool freely afterward.
What audio formats can I upload?
You can upload common audio formats including MP3, WAV, M4A, WebM, OGG, and FLAC. The file is decoded in your browser to a 16kHz mono waveform, which is the input format Whisper expects, and then transcribed locally. Microphone recordings are captured as WebM audio and decoded the same way. As with everything else, the audio is processed entirely on your device and never uploaded.
Q&A SESSION
Got a quick technical question?
Skip the back-and-forth. Get a direct answer from an experienced engineer.