AI audio editing

The foundation of text-based editing is transcription, the same speech-to-text technology you explored in Thing 10. The tool listens to your audio, creates a transcript, and maintains a precise link between every word in the transcript and the corresponding moment in the audio.

When you delete a word from the transcript, the tool removes the matching audio. When you rearrange sentences in the text, the audio rearranges to match. The transcript becomes your editing interface, and the waveform becomes something you can largely ignore.

If you work with spoken-word content, this is a big deal. Podcasters, trainers, anyone who records presentations or instructions: they all spend significant time editing out mistakes, hesitations, and tangents. Text-based editing cuts that work down because you can read what was said and make decisions at the speed of reading rather than the speed of listening.

It isn't perfect. Edits between words can sometimes sound slightly unnatural: a tiny gap or an abrupt transition that a careful listener might notice. The more natural and consistent your speaking pace, the cleaner the edits tend to be. But for most purposes, the results are surprisingly smooth.

Audio enhancement is a different kind of AI trick. Rather than editing what was said, it improves how it sounds.

Traditional audio cleanup requires someone who understands equalisers, noise gates, compressors, and de-essers. These are specialist tools that take time to learn. AI enhancement replaces all of that with a single step: upload your file, let the AI process it, and download the cleaned-up version.

Under the hood, these tools use machine learning models trained on thousands of hours of audio. They've learned to distinguish between human speech and everything else: traffic noise, air conditioning hum, keyboard clicks, room echo, the dog barking in the background. The AI isolates the speech and removes or reduces everything else, then applies subtle processing to make the voice sound clearer and more present.

A recording made on your phone in a noisy café can come out sounding like you were sitting in a quiet studio. There are limits, of course. If the background noise is louder than your voice, or if the recording quality is very poor, even AI can only do so much. But for the kinds of recordings most people actually make (phone recordings, laptop microphones, video calls), the improvement is dramatic.

Descript (descript.com) is the leading tool for text-based audio and video editing, with around six million users. It popularised the idea of editing audio by editing text, and it's still the best implementation of the concept.

When you upload audio to Descript, it transcribes the recording and presents you with a text editor that looks and feels like a word processor. You can select text and delete it, and the corresponding audio disappears. You can highlight filler words ("um," "uh," "you know," "like") and remove them all in one click. You can rearrange sections by cutting and pasting text, just as you would in a document.

Descript also includes Studio Sound, an AI enhancement feature that cleans up audio quality by removing background noise, reducing echo, and making voices sound clearer. On paid plans, there's also Overdub, which lets you create a clone of your voice that can speak new words, useful for correcting small mistakes without re-recording.

The free plan gives you 60 media minutes per month and 100 one-time AI credits. That's enough to try the core features, but the AI credits are consumed each time you use features like Studio Sound (around 10 credits per use), so you'll want to be deliberate about what you enhance. The free plan also watermarks exported video, though this doesn't affect audio-only exports. Paid plans start from around £13/month (Hobbyist) if you need more capacity.

Descript requires creating an account and downloading the desktop application (available for Windows and macOS). It doesn't run in a browser.

Adobe Podcast Enhance Speech (podcast.adobe.com/enhance) takes a different approach. It doesn't do text-based editing at all. It focuses entirely on audio enhancement. You upload a recording, the AI processes it, and you download a cleaner version. That's it.

What makes it worth using is how well it works and how little effort it requires. There are no settings to configure, no sliders to adjust (unless you're on the premium plan), and no learning curve. You upload, you wait, you download. The enhancement is often dramatic: background noise vanishes, echo reduces significantly, and voices become noticeably clearer.

The free tier requires an Adobe account (free to create) and lets you process files up to 30 minutes long, with a daily limit of one hour of total processing. That's plenty for experimentation. Files up to 500MB are supported, and it accepts common audio formats including MP3 and WAV.

There's a V2 version of the enhancement engine, which Adobe says produces more natural results. The free tier uses automatic enhancement, while the premium plan (£8/month, or included with Creative Cloud subscriptions) adds adjustable enhancement strength and support for video files.

One thing worth knowing: the enhancement is designed specifically for speech. It works well on spoken-word recordings but isn't designed for music or complex audio with multiple sound sources.

Auphonic (auphonic.com) offers automated audio post-production (levelling, noise reduction, and loudness normalisation) with a free tier of two hours per month. It's popular with podcasters who want a "set and forget" approach to making their audio sound consistent.

Audacity (audacityteam.org) is the long-standing free, open-source audio editor. It uses traditional waveform editing rather than text-based editing, but it now supports AI-powered noise removal through built-in tools and third-party plugins. If you want maximum control over your audio editing without paying anything, Audacity is worth knowing about, though the learning curve is steeper.

Krisp (krisp.ai) works differently. It removes background noise in real time during calls and recordings, rather than after the fact. If you frequently take calls from noisy environments, it's worth a look.

When you remove words or sentences in Descript, listen carefully to the join points. Sometimes the edit is seamless; sometimes there's a tiny click, an unnatural pause, or an abrupt change in room tone. The AI is very good at making clean cuts, but it isn't perfect, especially if your speaking rhythm changed between the words on either side of the edit.

Adobe's Enhance Speech tool is impressive, but it can sometimes make voices sound slightly processed or artificial, particularly if it's working hard to remove a lot of noise. Listen for any "watery" or metallic qualities in the enhanced version. This is a common trade-off in AI audio processing: the more aggressively the tool cleans up the audio, the more risk there is of introducing subtle artefacts.

Neither tool will produce broadcast-quality results from a truly terrible recording. But both can take a recording that you'd be embarrassed to share and turn it into something perfectly presentable. For most professional uses (internal presentations, training recordings, social media content), "good enough" is exactly what you need, and these tools get you there quickly.