Private browser production workflow
Turn one uneven spoken-word recording into a tighter, level-controlled and delivery-ready episode—without sending the file to a cloud service or bouncing between three separate tools.
Your recording stays inside this browser tab.
DYNAMICS MAP
Follow the level decision, not just the waveform. The important story is how the detector reacts, where gain starts moving and whether the result stays controlled without flattening the performance.
Solo podcasts, interviews, narration, lessons, audiobooks and voice-led video drafts that need a repeatable finishing pass without uploading the recording to a third-party service.
A 20 ms RMS envelope finds qualifying pause interiors, a linked soft-knee compressor applies transparent level control, and a gated K-weighted loudness estimate drives target gain with sample-peak protection when required.
Source and enhanced auditions, aligned dynamics Difference and removed-pause previews when memory permits, WAV/MP3 exports and a JSON receipt with timing, gain reduction, LUFS, peak and exact output dimensions.
This is deterministic DSP rather than generative voice restoration. Pause detection can catch quiet syllables, compression can raise noise, sample peak is not true peak and the loudness estimate does not replace a publisher-required certified meter.
Quick answer: AudioWrench Podcast Audio Enhancer is a free local production chain for spoken-word recordings. It shortens selected long pauses, applies conventional speech compression, estimates integrated loudness, moves the episode toward a chosen delivery target and gives you source/result auditions plus a processing receipt. It does not upload the recording or pretend to reconstruct speech with AI.
A 20 ms RMS envelope identifies sustained low-level regions. The editor removes only the middle of a qualifying pause and leaves an adjustable quiet edge, reducing the breathless cadence caused by deleting every silent frame.
A linked detector follows the loudest channel, smooths changes with attack and release timing and applies a soft-knee gain curve. Threshold, ratio and makeup gain stay visible instead of hiding behind an unexplained enhancement percentage.
The browser renders a K-weighting filter chain and uses 400 ms gated blocks for a BS.1770-style integrated-loudness estimate. Gain targets −16, −14 or −23 LUFS, with sample-peak protection when the projected result would exceed −1 dBFS.
Measurement boundary: integrated loudness is an in-browser estimate and peak is sample-based, not an oversampled true-peak measurement. Use the receipt for repeatability, then verify a final commercial delivery with the meter and specification required by your publisher.
Tighten long thinking gaps while preserving short expressive pauses, then reduce large level swings between quiet sentences and emphatic phrases.
Use a slower release and stronger ratio as a practical starting point for changing mic distance. The workflow does not remove cross-talk or match two separate speakers independently.
Prepare a voice-led render near −14 LUFS when that matches the next video workflow, then export WAV for the edit or MP3 for a review copy.
Remove obvious dead air, record the exact settings in JSON and retain the original source so later editorial decisions remain reversible.
Use AudioWrench when local privacy, explicit controls and repeatable delivery matter. Consider specialist restoration when a recording has severe reverb, broadband noise, clipped words or missing speech information. A final listening pass is required in either case.
This workflow is not speech synthesis, identity synthesis, transcription, speaker separation or generative restoration. It will not identify filler words, repair clipped consonants, remove every fan or room reflection, balance multiple isolated speakers independently or guarantee approval by a broadcaster. Pause detection can confuse quiet speech with room tone; compression can emphasize noise; a loudness target can still sound wrong when tonal balance or distortion is poor. The interface makes those boundaries visible so an editor can make an informed decision.
Quick answer: A guided local spoken-word production chain that shortens long pauses, controls speech dynamics, estimates integrated loudness, moves the result toward a delivery target and exports the finished episode.
Know before you use it: This is deterministic DSP rather than generative voice restoration. Pause detection can catch quiet syllables, compression can raise noise, sample peak is not true peak and the loudness estimate does not replace a publisher-required certified meter.
Privacy: The selected recording is decoded, analyzed, processed and encoded in the current browser tab. It is not uploaded to AudioWrench; normal page assets can still be requested as described in the privacy notice. Privacy details →
It can shorten detected long quiet spans, apply deterministic linked speech compression and apply measured gain toward a selected integrated-loudness target. Each stage is optional and the final receipt records what ran.
No. The selected audio is decoded, analyzed, processed and encoded in the current browser tab. Normal website assets can still be requested as described in the privacy notice, but the selected recording is not sent to AudioWrench.
−16 LUFS is a common spoken-word starting point, while some video and streaming workflows use −14 LUFS. A distributor can apply its own playback normalization, so check its current delivery specification and listen after encoding.
No. It uses transparent signal processing and does not synthesize missing speech, remove every noise source, clone a speaker or infer words. Damaged recordings can still need specialist restoration.
It is the final processed signal minus the pause-tightened pre-dynamics timeline. Pause removal changes duration, so timing edits are excluded from this subtraction to keep the compared buffers aligned.
After decoding, the tool estimates working memory from the real frame count, channel count and browser device signals. Lean mode keeps source and finished previews but skips extra difference buffers, reducing memory without enforcing a fixed input-file cap.
Load one recording, review the three stages and keep the final decision in your ears.