Free speech to text for interviews and podcasts, no upload

Transcribing a one-hour interview or podcast episode by hand takes three to four hours if you do it carefully, and it is exactly the kind of work that gets skipped when a deadline is close. The free speech to text tool on zalt.me runs OpenAI's Whisper model directly in your browser, no upload, no signup, no API key, and turns that hour of audio into a searchable, timestamped transcript in a couple of minutes on a normal laptop. It covers 99 languages with automatic detection, gives you both segment and word-level timestamps so you can jump straight to a quote, and can translate the result to English if the interview was not conducted in it.

I'm Mahmoud Zalt, an independent senior AI systems architect who has been building production software since 2010, that's 16 years, and I founded Sista AI (sistava.com), where autonomous AI agents run in production, not demos. I built this transcription tool, and the rest of the free tools collection, because journalists, researchers, and podcasters kept telling me the same thing: they didn't want to upload a recorded conversation, sometimes an off-the-record one, to a third-party server just to get a transcript.

Why interview and podcast transcription is its own problem

Most "how to transcribe audio" advice assumes a short, clean clip: a voice memo, a two-minute note. Interviews and podcasts are a different animal, and the friction shows up in specific places.

• Length. A single episode or sit-down interview usually runs 30 to 90 minutes, sometimes longer. Anything that is not close to instant becomes a real time cost across a season or a beat.
• Multiple speakers. Two or three voices trading turns, sometimes talking over each other, is harder for any speech model than one person reading into a microphone.
• Inconsistent audio quality. Remote guests on a phone line, a laptop mic in a noisy cafe, a Zoom call with echo. Real interviews rarely sound like a studio recording.
• The need to find things fast. You are not reading the transcript for pleasure. You are hunting for the one quote that makes the article, or the three minutes worth clipping for a highlight reel, and scrubbing through raw audio to find it wastes more time than the transcription itself.

None of that is solved by a generic "upload your file, get text back" tutorial. It needs a workflow built around noise, multiple voices, and timestamps, which is what the rest of this guide covers.

A practical workflow for a full episode

This is the sequence I'd actually use on a real interview or podcast recording, in order.

1. Clean up noisy audio first

If the recording has hiss, hum, fan noise, or a rough phone line, run it through the Audio Noise Reducer before transcribing. It uses RNNoise to strip background noise while leaving speech intact, and cleaner input audio means fewer transcription errors, especially on the quieter of two speakers. Skip this step if your recording is already reasonably clean, it is not always necessary.

2. Transcribe with Speech to Text

Open the speech to text tool, upload the MP3, WAV, or M4A file (or record straight from your microphone for a live conversation), and let Whisper run locally in the browser. There is nothing to configure beyond picking a language or leaving it on automatic detection. The audio never leaves your device, which matters when you are sitting on an unreleased interview, an embargoed story, or a conversation a source asked you to keep off any third-party server.

3. Skim with segment timestamps, then narrow with word-level ones

Once the transcript is ready, use the segment timestamps to skim the conversation in chunks and find the general area where a topic comes up. Then switch to word-level timestamps to pinpoint the exact start and end of the sentence you want to quote or clip. This two-step scan is faster than reading the whole thing top to bottom.

Timestamps are the whole point for this use case

For a voice memo, a plain transcript is enough. For an interview or podcast, timestamps are what make the transcript actually useful, because your real task is almost never "read the text." It's "find the 12 seconds that matter."

Journalists use this to lift an accurate, attributable quote without replaying the same thirty seconds of audio five times. Podcasters use it to build a chapter list or find the sound bite worth clipping for social. Researchers use it to locate exactly where a subject answered a specific interview question across a long session. In every case, the timestamp is what turns a wall of text into something you can navigate, instead of a transcript you have to read start to finish just to find one line.

Interview in another language? Translate it to English

A lot of interview and podcast transcription tools quietly assume everything is in English. This one doesn't. The underlying Whisper model supports 99 languages with automatic detection, so an interview conducted in Arabic, Spanish, French, or nearly anything else gets transcribed in its original language without you having to set anything manually.

If you need the transcript in English, whether for an English-language publication, a research report, or just your own notes, there's a translate-to-English option built into the same tool. It is not a substitute for a professional human translation on anything sensitive or legally significant, but for pulling quotes, writing a summary, or getting the gist of a foreign-language interview fast, it removes a step that used to require a separate service entirely.

What to do with the transcript once you have it

The transcript itself is rarely the deliverable. It's raw material. A few common next steps:

• Pull quotes for an article. Search the transcript text for a keyword, jump to the timestamp, verify the exact wording against the audio, and quote it with confidence.
• Write show notes. Skim the segment timestamps to build a rough topic list with timecodes, the same structure most podcast show notes use.
• Get a fast summary of a long interview. If you recorded ninety minutes and need the five-minute version for yourself or an editor, feed the transcript into Voice Notes, which pairs Whisper with a summarizer built for exactly this.
• Generate subtitles for a video episode. If the podcast is also published as video, the Subtitle Generator uses the same Whisper model to output SRT or VTT files ready to attach to the upload.

All three of these tools are part of the same free tools collection, and they're built to chain together: clean the audio, transcribe it, then summarize or subtitle from the result.

Where this breaks down, honestly

It's a genuinely useful tool, not a magic one. A few honest limits worth knowing before you rely on it for something important.

• Heavy cross-talk hurts accuracy. When two people talk over each other, or a third voice cuts in mid-sentence, the transcription gets noticeably less reliable in that exact stretch. Slow down and verify anything you plan to quote from an overlapping section.
• It doesn't label who said what. The transcript gives you the words and their timing, not automatic speaker names. On a two- or three-person interview you'll want to skim and tag speakers yourself, or listen at the timestamp to confirm who's talking.
• Long files take longer, locally. Because everything runs on your device instead of a server, a two-hour episode will take noticeably longer to process than a ten-minute clip, and the exact time depends on your machine's hardware.
• Strong accents, mumbling, or very poor audio still reduce quality. Whisper is good, not infallible. If the source audio is genuinely bad, no transcription model fixes that entirely, cleaning it up first helps but has limits.

Frequently Asked Questions

Can this transcribe a podcast episode with multiple speakers?

Yes. It handles multi-speaker conversations, including interviews and panel-style podcasts. Accuracy stays high during normal back-and-forth dialogue and drops somewhat during overlapping cross-talk, which is true of every speech recognition model, not just this one.

How long does it take to transcribe a one-hour interview?

Typically a few minutes on a normal laptop, since the Whisper model runs locally via WebAssembly rather than waiting on a server queue. Processing time scales with file length and your device's hardware, so a two-hour episode takes longer than a twenty-minute one.

Can I transcribe an interview that wasn't conducted in English?

Yes, the tool supports 99 languages with automatic language detection, and it can translate the transcript to English if you need it in that language for publication or notes.

Is my audio uploaded anywhere?

No. Everything runs locally in your browser using Hugging Face Transformers.js, there's no server call and no signup. The recording never leaves your device, which matters for off-the-record or embargoed material.

What if my recording has a lot of background noise?

Run it through the Audio Noise Reducer first. It strips hiss, hum, and background sound while preserving speech, which usually improves transcription accuracy on a rough recording before you transcribe it.

Get the quote, not just the recording

An interview or podcast recording is only useful once you can find what's in it fast. Clean the audio if it needs it, transcribe it, and use the timestamps to get to the exact quote or clip in seconds instead of scrubbing through raw audio.

Transcribe audio free -> or explore the other 63 free tools, including the noise reducer, subtitle generator, and voice notes summarizer that pair with it.