TTS Audio Quality Guide: Loudness, EQ & Export Settings

Start with the delivery specification

Before touching a processor, write down where the narration will be used: a video timeline, podcast feed, course, screen reader prompt, or a downloadable voiceover. Record the requested file type, sample rate, channel layout, loudness target, peak ceiling, and file-size limit. These requirements belong to the destination, so a preset that works for one project can be wrong for another.

Keep the original render as a clean source. If you are editing against video, match the session to the production timeline. For audio-only work, follow the delivery service or editor's accepted settings. Converting to a higher sample rate does not restore detail that was not present in the source; resample only when the handoff requires it. Keep a lossless working copy so you can make a later revision without encoding an already compressed file again.

Measure loudness and true peak

Integrated loudness describes the level of a complete programme, while short-term readings help reveal uneven sections. ITU-R BS.1770 specifies algorithms for measuring programme loudness and true-peak level. It is a measurement recommendation, not one loudness target for every kind of speech. For example, EBU R 128 recommends −23 LUFS for broadcast programme loudness; that broadcast value should not be copied automatically into a web, podcast, or app workflow.

Set the target from the actual brief. If there is no delivery target, preserve natural dynamics and avoid clipping instead of forcing every file to the same number. A sample-peak meter only checks the stored samples; the reconstructed waveform can peak between them. Check true peak as well, especially after lossy encoding. A limiter can control occasional overs, but heavy gain reduction often makes a synthetic voice sound harsh or tiring. If the measured level and the listening result disagree, inspect silence, phrase balance, and the source before adding more compression.

Process only an audible problem

Start by listening to the unprocessed render at a comfortable level. Text-to-speech voices are already produced sounds, so a generic mastering chain can remove useful detail as easily as it can improve clarity. Use equalization only when you can identify a problem such as low-frequency rumble, a nasal resonance, or excess brightness. Make a small change, bypass it, and compare at matched loudness; a louder version can seem better even when it is less clear.

Use a de-esser when repeated “s” or “sh” sounds distract from the words. Adjust its detection range to the selected voice and language, then listen for consonants becoming dull or lisped. A high-pass filter can remove unwanted rumble from a mixed recording, but it is not a required step for every generated voice. If the source is clean, leaving it unchanged is a valid engineering choice.

Export, then review the encoded file

Choose WAV when another editor needs an uncompressed handoff. Choose a lossy format only after checking the playback devices, upload service, and size limit for the project. Codec, bitrate, and sample-rate choices should follow that brief rather than a universal “web audio” preset. Keep the filename and version tied to the script and voice so the approved render can be reproduced.

Review the exported file, not just the editor session. Encoding can change peak readings, and an export can introduce a clipped start, a cut-off final consonant, or an unexpected gap. Listen once on headphones and once on a typical phone or laptop speaker. Check the first and last words, pauses, proper names, and transitions between segments. Then measure the final file again and compare it with the delivery requirements.

A repeatable review pass

1. Save the untouched render and write down the delivery requirements.
2. Listen for pronunciation, pacing, noise, and obvious level jumps before processing.
3. Measure integrated loudness and true peak with a meter that follows the selected delivery workflow.
4. Make the smallest correction that solves an audible problem; level-match before comparing.
5. Export, re-measure the encoded file, and listen on the target playback devices.
6. Keep the approved source, settings, and final export together for future edits.

This process makes decisions traceable without pretending that one loudness number, equalizer curve, or codec setting fits every voice. For browser-generated narration, it also separates model output from mastering decisions: first confirm that the words and timing are right, then shape the file for its destination.

Final checklist

• The selected file format and sample rate match the delivery brief.
• Loudness follows the destination target, and true peak is checked after encoding.
• EQ, filtering, compression, and de-essing solve problems you can hear.
• The final export has been checked for clipped starts, cut-off words, and level jumps.
• A clean source render and the approved export are both retained.

Author: Kokoro Web Team • Last updated 2026-10-15