saturalabs

Repair or Retake? Listen to Two Reproducible Dialogue Cleanup Tests

The same short passage was mixed with a steady low hum and with intermittent clicks. Hear the original and two actual FFmpeg outputs on each case, then compare them with the clean source. The intermittent-click case also includes one text-prompted speech estimate and what remains after separation. Compare both with the clean reference before judging the result.

Downloads

Listen to the files

Measured changes in the supplied files

Local filters in FFmpeg 8.0.1 and one SAM Audio Large run · 48 kHz, 16-bit mono WAV
SampleMethodPause RMS (dBFS)Difference from clean speech (dBFS RMS)
Steady humOriginal-23.90-23.91
Steady hum100 Hz high-pass-30.60-30.29
Steady humFFT denoise-24.06-23.22
Intermittent clicksOriginal-37.98-39.85
Intermittent clicks100 Hz high-pass-38.00-37.64
Intermittent clicksFFT denoise-38.92-30.35
Intermittent clicksSAM Audio speech estimate-74.94-40.16

Run your own listening check

  1. Preview the original, clean reference and repair tries on the same speakers at a comfortable playback level. If one sounds much louder, match playback levels in your editor before judging clarity.
  2. Start without the transcript. Write down the words you hear and mark any unclear ending or distracting artifact with a timestamp. Check the transcript and listen again.
  3. Compare the same moments. Mark when quieter pauses help and when a change to consonants, timing or tone would justify another take.

What the sample comparison measured

The pause measure is the recorded RMS level from 0.15 to 0.80 seconds. The click case deliberately includes a click in that window. The speech-window measure is the RMS of the sample-by-sample difference from the clean reference at the supplied times.

A larger difference can reflect noise, timing, phase, or changes to speech. Use your own listening pass to judge intelligibility. The clean reference comes straight from the synthetic speech generator before noise was added.

The lines are “The blue marker is on the desk,” “Please check the window before you leave,” and “Save the last clean take.” Speech was generated from these original lines with eSpeak NG. No third-party recording was used.

How the prompt-based comparison was run

We sent the eight-second intermittent-click mix to the pinned geopti/sam-audio-large model through Replicate with the prompt “The speaking voice.” The request asked for the residual and used no time anchors. The returned WAVs are provided as received. The JSON records the full model revision, prompt, settings, and file hashes.

Only the speech estimate is included in the measurement table. All files were compared from their first sample at their original gain. Listen around 1.53, 3.99, 4.24, and 6.36 seconds for residual clicks and clipped consonants. Listen to the remainder for missing bits of speech.

This is one run on generated speech with a particular mixture and prompt. Evaluate a representative passage of your own recording at matched playback levels before making an editing decision.

Reproduce the two local filters

To reproduce the second case, replace steady-hum in the names with interrupting-clicks. Use the clean reference and measured windows in the JSON report.

On a real recording, use small reversible changes. If you cannot verify an important word against the original or an alternate track, consider a retake when one is available.

ffmpeg -i steady-hum-original.wav -af 'highpass=f=100' -c:a pcm_s16le steady-hum-high-pass.wav

ffmpeg -i steady-hum-original.wav -af 'afftdn=nr=12:nf=-42:tn=0' -c:a pcm_s16le steady-hum-fft-denoise.wav

Continue the workflow