Voiceover, audio repair and mastering

What this actually solves

There is one asymmetry at the centre of this work and it explains every enquiry that arrives. Viewers forgive bad picture and never forgive bad sound. Footage shot on a phone in poor light gets watched. A recording with an air conditioner droning under it gets closed within seconds, no matter how good the content is.

So this work arrives in one of three shapes. In the first, you have excellent material and one fatal problem in the audio — an interview recorded in a room with hard walls, a lapel mic rubbing against a collar, a hum that was inaudible on set and impossible to ignore in the edit. You are not buying polish. You are buying the difference between a usable project and a wasted shoot.

In the second, the problem is consistency across a body of work. A course of lessons, recorded over months, in different rooms, at different distances from the mic, at different levels. Individually each one is acceptable. Played in sequence they are exhausting, and the exhaustion is what makes people stop taking the course.

In the third you simply need a voice — a script that has to be read well, in the right tone, at the right pace, ready to drop into an edit.

How the work runs

1. Assessment, before anything is promised. Every file listened to, and a plain statement of what is repairable and what is not. Steady noise, yes. Room reflections, partially. Clipped distortion, no. This half-hour is what prevents a delivery-day argument. One to two hours.
2. Capture or generation. If a voice is being produced rather than repaired - the script is prepared for reading first, with long sentences broken, breath points marked, and terms like SaaS, ROI and numbers written the way they are spoken. Two or three sample reads go to you before the full script is committed. Half a day to two days.
3. The repair and tone chain, in a fixed order. Gentle noise reduction, then EQ — clearing the rumble below roughly 100 Hz that carries no speech, with a modest lift around 3 to 5 kHz where consonants live and intelligibility comes from. Then de-essing, then compression to bring the loud and quiet passages into a usable range. One to four days depending on how many files.
4. Consistency and loudness across the set. Files matched to each other rather than each perfected alone, silences filled with the room's own quiet rather than digital nothing — true silence between sentences reads to a listener as a dropout — and the whole set normalised to its target. Half a day to a day.
5. Delivery. A WAV master, MP3 distribution files, and where it is going back into video, a clean voice stem with nothing else baked in. A few hours.

Worked in Reaper for anything serious, with Adobe Podcast Enhance used selectively — it can rescue a difficult recording and it can also introduce a synthetic, processed quality, so its output is compared against the original by ear every time rather than trusted by default. ElevenLabs for AI reads, and Auphonic for loudness where a set needs matching quickly.

What moves the price

The range is $50–$1,500 and turnaround is 2–10 days, which is a wide spread for a service where the individual steps barely change. Three things account for it.

How damaged the source is. A clean recording that needs levelling and mastering is a fast job. A recording with hum, room echo, mic handling noise and inconsistent distance is a slow one, because each of those needs a different treatment and they interact — fixing one can expose another.

How many files there are. A single sixty second read or one podcast episode sits at the bottom of the range. A complete course of lessons, matched to each other and mastered as a set, sits at the top.

Whether a voice is being produced or only processed. Producing a read means script preparation, casting or voice selection, sample rounds and re-records when the script changes. Processing existing audio starts with the material already in hand.

What usually goes wrong

Noise reduction is turned all the way up. It is the most tempting slider in audio and the most destructive. Push it far enough and the hiss disappears completely, along with the natural top end of the voice — the result sounds underwater, and the s sounds turn into something metallic and strange. The discipline is to reduce noise until it stops being distracting and then stop, because a small amount of remaining room noise is far less noticeable than a voice that has been hollowed out.

Someone promises to repair clipped audio. When a recording hits the digital ceiling, the tops of the waveform were never captured. There is nothing there to restore. Tools can disguise the harshness, and the word still sounds broken. Committing to fix it and discovering this at delivery costs you your re-record window, which is why it is said on the first day.

Each file is treated in isolation when it belongs to a set. Lessons individually optimised will still not match, because each was optimised against its own flaws rather than against the others. The chain is built once, against the weakest file, applied across everything, then checked in sequence — which is a different job from making any single file sound its best.