# Can’t Listen to a Voice Note? Turn It into Editable Text with Whisper

> Source: <https://discuss.huggingface.co/t/can-t-listen-to-a-voice-note-turn-it-into-editable-text-with-whisper/192027#post_1>
> Published: 2026-11-18 09:00:41+00:00

**Recommendation:** Keep Whisper's raw transcript alongside your edits and check names, numbers and negations against the audio before relying on the text; the sample run misheard "for a fortnight" as "of refort tonight."
More in the playbook "Verifying model output": <https://wpnews.pro/playbooks/output-verification>

A friend sends a voice note while you are somewhere quiet. You want to find the useful sentence without playing the whole message aloud. A text draft can help, as long as you can check the words that matter later.

This small Gradio app turns an uploaded recording into editable text using [Whisper tiny](https://huggingface.co/openai/whisper-tiny). It keeps the original model output beside your edited version, provides audio playback, and saves the edited text as a TXT file. It runs locally; there is no public Space to open yet.

The walkthrough uses a permitted public reading, not a private message from a friend. The screenshots come from the actual running Gradio app and real model inference.

The accompanying project contains the app, a pinned dependency list and a setup README. Follow that README with Python 3.12, then start the app and open the local app on the same computer. The first model download needs an internet connection. After loading it, the app performs transcription on the local CPU rather than calling a hosted inference API.

For a first try, choose **Load public LibriSpeech sample**. You can also upload the bundled WAV yourself, which is how the browser test was performed. It is a short English reading supplied through the official Hugging Face model page.

This version accepts **mono WAV audio at 16 kHz, up to 30 seconds and smaller than 10 MB**. If your phone exports M4A or MP3, this app does not directly support that file. Export a suitable WAV from your own audio editor; changing a filename extension is not conversion.

For your own recording, first make sure you have permission to transcribe it. The app accepts a file you deliberately choose; it does not read messages inside WhatsApp or another chat app.

Press **Transcribe**. The Original model transcript box fills with Whisper’s output. The editable box below starts with the same text, so there is something ready to work on.

Real local Gradio inference on a public LibriSpeech reading. The lower box initially copies the original output; this is not an online Space.

In this actual sample run, the original included “day after day of refort tonight.” It looks like a sentence, but the ending is not useful English in this context. That is a good reason to keep the recording nearby rather than treating a tidy-looking transcript as checked text.

Do not press Transcribe again just to change punctuation. Type in the lower box instead. Another inference would replace your working copy with a new model result.

When you can listen, use the original audio player and follow the text. Check a name, a time, an address or a negative such as “not” before relying on it. If the recording is unclear, ask the sender rather than completing the sentence from a guess.

For this walkthrough, the editable phrase was changed from “of refort tonight” to “for a fortnight.” That matches the wording in Hugging Face’s published example for the same audio. The change was entered into the editable box; it was not an automatic correction by the app.

The phrase was edited in the real text box, using the published sample wording as a reference. The app did not make the correction automatically.

The original box still shows the mistaken phrase. Having both versions makes it easy to see which words you changed. The app does not highlight uncertain words or check your correction for you. With a real voice note, you usually will not have a published reference to consult.

For a short note, use the copy icon in the editable box. For a file you can keep, press **Save edited TXT**, then click the download that appears below it.

Actual TXT export from the edited box. The downloaded contents were checked against that text; no public Space was deployed.

The downloaded file contains the current edited text, not the read-only original. The browser check downloaded it and compared its contents with the edited box; they matched, with a final newline added. Blank text is rejected instead of producing an empty download.

Upload a different recording to begin another note. The interface clears the previous transcript and download so that an old result is not left beside a new audio file.

The project ran the multilingual Whisper tiny checkpoint on a Mac CPU and tested this English sample. It also rejected a missing file, the wrong sample rate and a recording longer than its limit. That demonstrates this local workflow; it is not an accuracy or speed promise for other recordings, languages or devices. The model card describes recognition limitations, including text that was not spoken.

The audio is **LibriSpeech / LibriVox**, hosted by Hugging Face, under [CC BY 4.0](https://www.openslr.org/12/). It was converted from FLAC to WAV without changing its content; full attribution is included with the project.

This local app retains temporary uploads and exported files in its runtime folder. Stop the app and remove those files when you no longer need them. Once the draft is checked, keep the useful sentence—and the original recording until you are satisfied with the words.
