How Loud Is That Espresso Machine, Really?

10-second pump samples from 10 machines, side-by-side. With reference sounds.

About these recordings

Numbers like "62 dB" don't mean much in isolation. We extracted 10-second pump samples from authoritative YouTube reviews using yt-dlp + ffmpeg, then normalized loudness to -16 LUFS so all samples play at fair-comparison level.

Methodology details: how we recorded these →

Reference sounds

Compare with:
Quiet office · 50 dB
Conversation · 60 dB
Vacuum · 70 dB
Hair dryer · 80 dB

Espresso machine samples

Click ▶ to play. Other samples auto-pause for A/B comparison.

Cafelat Robot · ~0 dB · manual lever, near-silent · Review →
Flair 58 · ~0 dB · manual lever · Review →
La Pavoni Europiccola · ~0 dB · manual lever · Review →
Wacaco Picopresso · ~0 dB · hand-pumped · Review →
Breville Bambino Plus · 62 dB · ≈ conversation · Review →
Breville Barista Express · 65 dB · Review →
DeLonghi Stilosa · 68 dB · Review →
Lelit Anna · 69 dB · Review →
Gaggia Classic Pro · 70 dB · ≈ vacuum · Review →
Rancilio Silvia · 71 dB · Review →

How to test in your apartment

Real apartment noise depends on three variables not captured by 1ft dB measurements:

To estimate real-world impact: take the 1ft dB number, subtract 15 dB for distance to typical bedroom (15 ft), subtract 20 dB for one drywall barrier. So Bambino at 62 dB → about 27 dB through wall and bedroom — below sleep-arousal threshold.

Pipeline transparency

How we made these recordings

  1. Source identification — for each machine, we identified 2-3 authoritative YouTube reviews (Lance Hedrick, James Hoffmann, Whole Latte Love, Seattle Coffee Gear)
  2. yt-dlp download — full video downloaded for offline processing
  3. Whisper transcription — auto-generated timestamped transcript of the spoken audio
  4. Claude Haiku scan — AI reads transcript looking for "now we pull a shot", "here's how it sounds", "listen to this", and similar markers indicating the machine is running
  5. ffmpeg extraction — 10-second window starting at the identified timestamp, extracted to mp3
  6. Loudness normalization — ffmpeg loudnorm filter brings all samples to -16 LUFS (broadcast standard)
  7. Original video linking — every sample includes a link to the source video for verification

This pipeline runs locally via /scripts/2_extract_audio.py. Sandbox-extracted mp3 files are served from /assets/audio/ on the live site.