You recorded a meeting, webinar, interview, podcast, or voice memo. The words are there, but so is the laptop fan, HVAC rumble, keyboard clatter, or a roommate's blender. You need a transcript, not an audio restoration project. The good news: you do not need studio silence. You need speech that is intelligible to a transcription engine. The bad news: hitting one Enhance button on a clipped, echoing file usually makes transcription worse. This workflow focuses on cleanup that helps speech-to-text, with real settings and tradeoffs.
1. Diagnose the noise before you change anything
Listen with headphones in a quiet room. Play a 20-second section where nobody is talking. That noise-only sample tells you what you are fighting.
- Broadband hiss or fan: steady high-frequency noise. Noise reduction or AI cleanup works well.
- Tonal hum: low 50 Hz or 60 Hz buzz, often with harmonics. Use a de-hum filter, not broadband reduction.
- Intermittent clicks: keyboard, mouse, chair squeaks. Manual repair or a gentle gate works better than aggressive denoise.
- Reverb or clipping: cleanup tools cannot remove room echo well, and clipped peaks cannot be restored.
Open the file in Audacity or Ocenaudio and switch to spectrogram view. Hum appears as horizontal lines. Clicks appear as thin vertical spikes. Broadband noise looks like gray haze. This check stops you from applying the wrong fix.
2. Do levels, edits, and high-pass filtering first
Order matters. If you denoise before trimming, you waste processing on silence. If you normalize after aggressive processing, you amplify artifacts.
- Duplicate the original track. Never edit the only copy.
- Trim dead air and false starts, but leave 0.3 to 0.5 seconds around speech.
- Normalize peak level to -3 dB for filter headroom.
- Apply a high-pass filter at 80 Hz for most voices. Use 100 to 120 Hz if rumble remains. A 12 dB per octave slope is usually enough. Too high and voices sound thin.
- If you see steady low hum, apply de-hum at 50 Hz or 60 Hz plus harmonics. Reduce 10 to 20 dB; more can hollow out speech.
3. Apply noise reduction with conservative settings
For constant fan, AC, or hiss, broadband noise reduction is next. In Audacity, select 0.5 to 1 second of noise-only audio. Go to Effect > Noise Removal and Repair > Noise Reduction. Click Get Noise Profile. Then select the whole file and open the effect again.
- Noise reduction: 6 to 10 dB for mild noise, 10 to 12 dB for heavy fan noise. Above 15 dB you often get metallic artifacts.
- Sensitivity: 6 to 8. Higher catches more noise but can eat soft consonants.
- Frequency smoothing: 3. Bands 3 to 6 are common for speech.
- Preview a sentence with p, t, and k sounds. If they click or disappear, lower the reduction.
Listen for a robotic tail on words. That means you pushed too hard. Undo and try 2 to 3 dB less. Two gentle passes at 6 dB can beat one pass at 12 dB, but each pass adds artifacts.
4. Use AI noise removal for long or remote recordings
Remote meeting recordings have a predictable problem: every participant has a different room, mic, and internet compression. Manual cleanup on a 60-minute Zoom call is tedious. AI speech cleanup can remove HVAC, fan, and keyboard noise in one pass while preserving voice. Adobe Podcast Enhance, Krisp, Descript Studio Sound, and a free online option like ClearVoiceHQ (clearvoicehq.com) all target this use case.
AI cleanup is fast, but not magic. It can create watery artifacts, remove breaths, soften room tone, or change a voice's character. It also may not fix crosstalk, clipping, or heavy reverb. Use it after trimming and high-pass filtering, not instead of them. Keep the unprocessed original so you can compare transcripts and re-run with different settings.
5. A practical workflow for a noisy meeting file
- Archive the original video or audio file. Rename it with the word original.
- Extract a mono working file. With ffmpeg:
ffmpeg -i meeting.mp4 -vn -ac 1 -ar 16000 -c:a pcm_s16le meeting.wav. - Open the file and mark a noise-only section.
- High-pass at 80 Hz, then de-hum if needed.
- Run conservative noise reduction using the profile from step 3.
- If noise is still distracting, run one AI cleanup pass. Compare before and after on the same sentence.
- Manually repair clicks, coughs, and keyboard spikes. Use spectral repair or replace with room tone.
- Normalize to -3 dB peak, or -16 LUFS if publishing. For transcription, consistent level matters more than exact loudness.
- Export a clean copy as 16-bit PCM WAV, 16 kHz, mono. Most speech-to-text engines, including Whisper-based tools, work well with this format.
- Transcribe the cleaned file, then proofread while listening at 1.25x speed. Cleanup helps accuracy, but names, acronyms, and crosstalk still need human review.
6. Export settings and tradeoffs for transcription
- Sample rate: 16 kHz mono is enough for speech and keeps uploads small.
- Bit depth: 16-bit PCM is standard. Avoid low-bitrate MP3 before transcription; compression artifacts confuse some engines.
- Loudness: Peaks around -3 dB, average speech around -18 to -16 LUFS. Too quiet misses words; too loud clips.
- Silence removal: Removing long gaps can speed transcription, but leave 0.3 seconds around speech or you cut off word beginnings.
- Music beds: Mute or lower intro music under speech. Noise removal is not a music separator.
- Crosstalk: If two people talk at once, no cleanup tool will separate them perfectly. Transcribe first, then edit against the audio.
- Over-denoising: Robotic voices hurt transcription. Natural speech cues matter.
- Original file: Always keep it. You may need to re-process after hearing the first transcript.
7. When cleanup is not enough
If the signal-to-noise ratio is terrible, cleanup can only do so much. Try a transcription service that handles noisy audio well, such as Whisper large-v3, Rev, or Otter. Upload the cleanest WAV you can. If accuracy is still low, run automatic transcription and have a human correct it while listening. For interviews or webinars, ask the other party for their local recording. A clean local track often beats any processed mix.
The goal is not perfect audio. It is a transcript you can trust without spending three hours on noise removal. Diagnose the noise, trim and filter, use conservative denoise, add AI cleanup only if needed, and export a speech-friendly WAV. That sequence solves most noisy meeting recordings and gets you much closer to a ready transcript.